Purpose and foundations
Memory across interactions
| Representation | Who writes it? | What reaches a new session? | Where can someone correct it? |
|---|---|---|---|
| Early ChatGPT saved facts | The assistant extracts an item from a user statement. | Saved items are supplied to later conversations. | The user could inspect the list and delete a mistaken item. |
| Later ChatGPT profile in Khemani’s test | A background process synthesizes conversations. | The assistant receives a derived account of the user. | Khemani could identify a wrong assertion but could not inspect the raw profile in settings. |
| Early Claude conversation search | Earlier conversations supply the searchable messages. | The assistant can look up past exchanges by topic or time. | Khemani’s account of Anthropic’s implementation leaves the correction controls unspecified. |
| Holmes’s Markdown note | Author and enrichment skill. | The file is available for lookup. | In the file itself. |
At 5:41 in Holmes’s demonstration, compare the dictated Ferrari note with its tags, source URL, enrichment timestamp, and backlinks. The skill can reuse existing tags. Follow a backlink to see which note refers to this one, or check a tag against Holmes’s original wording.
An agent is software that uses observations to choose actions. For a stored note to inform a model’s choice, some representation of it must enter the input to that call. A checkpoint has a different purpose: it preserves enough state to resume an execution. An inference cache saves numerical work performed while processing inputs. Neither specifies which past decisions a new conversation should inherit.
The distinction matters when an assistant seems to know something it learned yesterday. It might have received a saved item or a compressed account of the conversation. As Jack Morris explains, compression works on information already supplied. Training can instead change the model’s parameters so it behaves differently without receiving that record again. Editing Holmes’s file changes what a later reader can retrieve. It requires no such training. Continual Learning
Episodes and reusable claims
An episodic record describes what happened on a particular occasion. A transcript preserves the exchange in order. A shorter episode might keep just the request, the action, and the observed result. A semantic record expresses a claim for reuse across occasions. If someone asks for bullet points in this report, the episode tells us how they wanted that report formatted. Saving prefers bullet points would turn a one-time instruction into a default. If the user says use bullet points for future reports on this project, that wording supports a standing preference and tells the next session where it applies.
A semantic record can hold a stated preference, a checked fact, or a tentative inference. Its purpose is reuse. Semantic search is a method for finding records by similarity of meaning and works on either type. Sometimes the occasion is the better unit to retrieve. In Diane Lin’s alert-review example, a previous analyst judged a similar security alert benign. A new investigator can compare the circumstances even before anyone can state a general rule. If the precedent leaves the verdict disputed, an expert can identify the deciding condition. Keeping the original case makes that condition easier to check on the next alert.
Origins and turning points
Discussions of language understanding helped prompt Endel Tulving’s 1972 distinction between episodic and semantic memory. You can know what a word means without remembering when or where you heard it. In software the distinction describes what a record preserves. A system can look up a preference directly, but to explain how it learned that preference it needs a record of the occasion. Tulving’s retrospective
One early problem was how to learn enough about a person to help them without making every choice permanent. Elaine Rich’s 1979 GRUNDY book recommender retained individual reader profiles between sessions. At lookup, information specific to the dialogue took precedence over the persistent profile. A reader could ask for an unusual book without losing their usual preferences. Profile assertions also carried ratings and justifications. That made the reasoning behind a recommendation representable alongside the recommendation itself.
Another problem was finding an experience whose relevance became clear only later. Janet Kolodner’s 1983 paper on CYRUS describes answering questions about events involving Cyrus Vance and Edmund Muskie. A follow-up question could inherit clues from the previous exchange. Subject, time, and circumstances narrowed the events worth examining. In their 1994 case-based reasoning framework, Agnar Aamodt and Enric Plaza connected retrieval to testing. A system could reuse a previous solution, try it, revise it in light of the result, and select the useful parts for retention. Finding a precedent was the beginning of an investigation.
In the 2023 Generative Agents simulation, Joon Sung Park and colleagues stored observations and model-generated reflections that could inform later plans. The system could subsequently retrieve its own interpretation, making the distinction between observation and inference consequential. Charles Packer and colleagues’ MemGPT addressed a different constraint. The model could ask a function to search external storage, and the runtime would place a bounded page of results in the next input. The Letta workshop illustrates two write paths in this design. Conversation events populate searchable history automatically, while the agent deliberately deposits material in archival storage.
Forming accountable records
Identity, scope, and provenance
A message from a project lead saying that a reviewer wants email updates for this release describes the reviewer. An extractor that saves the preference under the lead’s personal profile has attached it to the wrong person. Even under the reviewer’s name, the claim needs its qualification: the lead reported it, and it applies to one release. Provenance records how the assertion came into existence, including the source, the transformation, and the actor responsible. It lets a reader inspect the basis for believing a claim. W3C provenance model
| Record detail | Meaning in the example |
|---|---|
| Claim ID, revision, content | The ID follows this assertion across edits. The revision identifies the wording a reader received. |
| Speaker, subject, recorder | The lead spoke about the reviewer. The extraction service wrote the record. |
| Attribution and evidence | Mark the preference as reported by the lead. Link the specific message version. A stronger inference needs its own justification. |
| Scope, owner, readers | Limit use to this release. Identify who controls the record and check who may read it. |
| Times and status | Record when it applies, when the system learned it, and whether it remains current, disputed, superseded, or restricted. |
What a source link says about a claim
ExampleThe original report and the evidence for revision have different roles.
Read the diagram as text
- Original source episode. Contains the report from which the assertion was extracted.
- Patient’s reported penicillin allergy. A derived relationship whose source and validity can be inspected.
- Later contrary episode, if present. Records the evidence responsible for invalidation.
- Original source episode → Patient’s reported penicillin allergy: source for extraction.
- Later contrary episode, if present → Patient’s reported penicillin allergy: source for invalidation.
Use stable identifiers for the identities the system can establish. For an authenticated account, OpenID Connect specifies the combination of issuer and subject. An email address can change hands. A stable login still leaves two separate decisions: resolving a person mentioned by the speaker and authorizing the requester to read the resulting claim. Mem0’s entity-scoping documentation makes a related distinction between the speaker identifier on an extracted record and entities mentioned in its content. People and actors · Origins and dependencies
A knowledge graph represents entities and their relationships explicitly. In Daniel Chalef’s Graphiti demonstration for Zep, a relationship records a reported penicillin allergy. As Chalef puts it, “Tracing a fact to its source is just a graph walk.” Look at the roles of the links in the figure. One episode produced an assertion. A later episode can explain its invalidation. Counting both as votes for the assertion would reverse the meaning of the second link. If two entity records merge, the combined identity needs the source links from both. Recording at 4:54
Choosing what to retain
Memory formation starts with a decision about future use. What later question would this information answer? How quickly might it change? Who would bear the cost of a mistaken assumption? In an exploratory study involving six interviews and 54 public discussion threads, people described wanting writing preferences to carry between conversations while keeping health information out of academic writing. Some wanted overlap between projects under their control. A single account-wide rule would obscure precisely the distinction they cared about. Users’ Expectations and Practices with Agent Memory
| Later need | What to keep | When to review or stop using it |
|---|---|---|
| Investigate a failed fix | An episode linking the attempted fix, observed result, and code version. | Recheck the assumptions against current code before reusing the fix. |
| Format reports within a project | The explicitly requested default, its project scope, and its source. | Honor a local exception. Review the default when the project closes. |
| Determine what is deployed now | A reference to the authoritative deployment record. | Read the live value for each current-status question. A dated snapshot serves a historical question. |
| No later task needs an incidental personal detail | No cross-session record. | Check that automatic extraction also leaves the detail out of summaries. |
An explicit request to remember gives evidence of intent for the stated use. Automatic extraction requires a policy about which observations qualify. Inferring a tendency introduces a further question about whether the evidence supports generalization. Across all three, a retention policy specifies purpose, readers, duration, review, and disposition, the action due at the end. A record can become useless before its permitted lifetime ends, or remain useful when deletion comes due. Tie the end action to an observable date or event. Delete the project default at closure gives a maintainer more direction than keep while useful. Artifact lifetimes · National Archives disposition guidance
From observation to accepted memory
Extraction proposes what a source means. Acceptance determines whether that proposal warrants changing a record. In Khemani’s account, conversations about choosing between Thailand and Turkey became a profile listing visits to both. He had traveled to Thailand and had never visited Turkey. The source supported considered Turkey. Saving visited Turkey changed the status of an event. A search that faithfully returns the saved profile will faithfully reproduce that mistake.
Use a direct mapping when an authoritative field already expresses the needed value and scope. For conversational text, compare the proposed record with its source. Check the speaker and subject, negation, time, and conditions. A plan needs evidence of completion before it becomes a completed event. For work performed through a tool, the tool result can supply that evidence where the agent’s own announcement cannot. If a permitted inference remains ambiguous and matters to a later decision, seek confirmation. Otherwise the application can retain the permitted episode without adopting the stronger claim.
The decision to write a reusable claim
ExampleA permitted source can remain useful even when the proposed interpretation fails.
Read the diagram as text
- Source passage. Includes attribution and circumstances.
- Proposed assertion. Identifies the subject and scope of the proposed change.
- Check meaning and permitted use. Inspect the evidence and the record the proposal would change.
- Commit the change.
- Seek confirmation. Keep the assertion out of use while clarification is pending.
- Leave the claim unwritten.
- Source passage → Proposed assertion: extraction.
- Proposed assertion → Check meaning and permitted use: candidate and source link.
- Check meaning and permitted use → Commit the change: supported and permitted.
- Check meaning and permitted use → Seek confirmation: permitted, materially ambiguous.
- Check meaning and permitted use → Leave the claim unwritten: unsupported or prohibited.
The Mem0 paper separates candidate extraction from updates. A new message pair, nearby messages, and a conversation summary supply candidates. A second stage retrieves related records and selects ADD, UPDATE, DELETE, or NOOP. This gives the application an identified change to inspect. A valid structured output tells the store how to interpret that operation. The evidence and retention policy determine whether to accept it.
A request-time write can make an accepted claim available before the application replies, at the cost of latency. A background writer can wait for more of the exchange, but the next session may arrive first. Be precise about acknowledgments. Receiving a message, committing an extracted claim, and publishing that claim to search are three observable events. An application that promises recall in the very next session needs a read path that can already see the committed revision.
Organizing and recalling memory
Consolidation and independent support
Consolidation reorganizes accumulated records for later use. It can merge duplicates or draw a lesson from several episodes. To judge that lesson, trace each claim back to its origin. In the figure, a test report says that reader A accepted an export format, and a project note quotes the result. Citing both documents still gives the agent one test. If the report is corrected, the quotation and any guidance based on it need review. The compatibility note about reader B has a different origin. It explains why the guidance for A should stay specific to A. Retrieving the guidance again adds no evidence for either claim.
Keep a derived lesson linked to its observations, qualifications, and conflicting evidence wherever source access and retention permit. A model-generated reflection interprets what happened. In Reflexion, an evaluator assesses an attempt and a model converts feedback into guidance for the next attempt. A mistaken explanation of a failure can persist even when the failure itself was observed correctly. Preserving the distinction allows a maintainer to withdraw the lesson while leaving an accurate episode intact.
Two documents can describe one test
ExampleFollowing derivation links reveals which apparent sources share an origin.
Read the diagram as text
- Test report for reader A. Reader A accepted the tested export.
- Excerpt in project notes.
- Compatibility note for reader B. Reader B requires a different format.
- Reuse this format for reader A. Guidance scoped to the tested reader.
- Test report for reader A → Excerpt in project notes: copied into.
- Test report for reader A → Reuse this format for reader A: supports reuse for A.
- Excerpt in project notes → Reuse this format for reader A: repeats the same test.
- Compatibility note for reader B → Reuse this format for reader A: excludes B from scope.
Anthropic announced dreaming for Claude Managed Agents as a research preview on May 19, 2026. Its Dreams documentation describes consolidating a memory store and past session transcripts into a separate output store. The original stays available for comparison. An output ID can appear while work continues, and cancellation can leave partial results. Check completion and inspect the proposed changes before adopting the new store. By contrast, context compaction shortens the input used for ongoing work. A consolidated store can influence many future inputs.
The same traces can also become training material. Samuel Denton distinguishes offline training on accumulated production traces from improvement during live interaction. Ronak Malde describes a form of distillation, training a student from a teacher’s signal, in which the student generates attempts and the teacher receives additional hints. Later behavior can change because the student’s parameters changed. With an external store, the useful lesson must instead be made available to the later decision. That difference determines whether there is an individual record a user can inspect and edit.
Records and searchable views
Give the accepted record a stable identity and identify the revision a later task reads. In a compact profile, correcting one sentence can require reviewing other sentences derived from the same assumption. Individually addressable claims let the maintainer change one assertion and follow its dependencies. An episodic log preserves the order surrounding an event. A linked note preserves wording a reader can check directly. For instance, Polygraph’s session debrief can orient a coding task with a summary, then direct the investigator to the logs, pull request, and current code to check whether an old assumption still holds.
An index is a representation built to find records. In OpenClaw’s built-in memory engine, files hold the retained text and SQLite indexes it for keyword search. Optional embeddings represent passages numerically for similarity search. A full rebuild prepares a replacement database while queries continue using the published one. The engine publishes the replacement atomically after a successful rebuild. This preserves search availability even when a rebuild fails, but a direct file read can see an edit before search does.
A corrected record has two read paths
ExampleSearch can serve older wording after the authoritative write has succeeded.
Read the diagram as text
- Memory file, revision 2. The correction is committed.
- Authorized direct lookup.
- Published index with revision 1.
- Replacement index with revision 2. Built separately from the serving index.
- Search result for the same record. Inspect the wording and revision returned.
- Memory file, revision 2 → Authorized direct lookup: current text by ID.
- Memory file, revision 2 → Replacement index with revision 2: revised text for indexing.
- Published index with revision 1 → Search result for the same record: old text until publication.
- Replacement index with revision 2 → Search result for the same record: new text after publication.
Check the promised behavior at the point where a new session reads. Zep’s ingestion guide distinguishes accepted input, completed processing, and search readiness. Its search_when_ready helper can finish as soon as any result appears. That result could predate the edit. For a required correction, verify the intended ID, revision, and content through the actual search path, or supply the authorized current record by ID. Where record changes feed an external index, a transactional outbox can commit the change and the obligation to index it together. Search visibility still awaits delivery and publication.
Recall for the current task
Recall selects earlier evidence for the decision at hand. Supplying a small profile on every call makes common defaults available immediately, but irrelevant entries take up space and can distract the model. Looking up records on demand costs time and lets the agent choose how much detail to retrieve. A saved claim might give the chosen export format in one line. To explain why that format works, the agent may need the original test report. The claim and the report are different retrieval units. Either can narrow the next investigation if it answers the question the agent would otherwise have to research again. Yu Su on search and expertise
| Candidate | Eligible? | Consequence |
|---|---|---|
| Another reviewer’s almost identical request | No. Wrong subject. | Textual similarity cannot transfer a preference between people. |
| An earlier, superseded default | No for current use. | It may qualify for a historical query. |
| A relevant report the requester cannot access | No. Access restricted. | Keep its contents out of the model input. |
| An accessible, current release preference phrased differently | Yes, within this release. | Supply the preference with the reporter and qualification intact. |
Use a known ID for exact lookup, date constraints for a past state, and similarity search for unfamiliar wording. Among eligible records, relevance, recency, and importance can help decide which to supply. Expand a compact claim to its supporting passage if the task turns on a missing qualifier. Check access on that expansion too. A scope filter enforces only its specified dimensions: Mem0 documents that a user-only filter can span applications and runs.
Deliver selected records with their versions, source roles, and applicable times to the context assembler. A completed lookup that finds no eligible record can justify asking for the missing information. An unavailable store or partially indexed import calls for a different explanation: the application could not complete the lookup. Treating either outcome as evidence that the user has no preference can turn a temporary read failure into another wrong memory. Degraded search coverage
Use and change
Personalization and authority
Personalization adapts behavior using information associated with a person or task. If someone has asked for bullet points in project updates, that standing preference can settle an open formatting choice. If today’s request specifies a narrative for the board, write the narrative. The standing preference can still apply to tomorrow’s ordinary update. An inferred tendency warrants a lighter touch. Repeated choices might all reflect the same temporary constraint, and silence gives little evidence about why someone accepted a result. Seek confirmation when the assumption would materially affect the outcome.
A correction is easier when the person can see both the claim and its basis. Cook and Kay’s 1994 Justified User Model exposed evidence for and against an inferred characteristic. A user edit added strongly weighted evidence from the user. The earlier justification remained inspectable. For a contemporary memory interface, this suggests a concrete distinction to expose when correcting a response: change the choice for this occasion, or change the default used on future occasions.
A source can be authoritative about one question and unable to settle another. A calendar can show that a time is free. Booking it requires current authorization. AuthMem-Bench tests this boundary by varying who supplied a claim while holding its content and the later task fixed. Stored text can also be hostile. Memory poisoning contaminates retained records so later retrieval can influence behavior. In AgentPoison, the attacker had access to an embedding model and could insert records. The attacker crafted malicious demonstrations that the retrieval system was likely to select when a later query contained a particular trigger. Familiarity with a retrieved instruction confers no permission to execute it. Current authority · AI Security
Correction and supersession
| What the evidence shows | How to change the memory |
|---|---|
| The extraction misstated the original. | Correct the assertion and review any summaries built from the error. |
| The old value was true until a real change. | Supersede it from the effective date. Preserve its earlier interval where retention permits. |
| The new request applies to one occasion. | Record or apply the narrower exception while keeping the broader default. |
| Two reports conflict about the same circumstances. | Keep both attributed accounts available for investigation. Seek evidence with authority over the particular question. |
| An inference has insufficient support. | Suspend its use. Ask for confirmation if the task depends on it. |
In XTDB’s temporal model, valid time records when a claim applies in the world. Recorded time, called system time in XTDB, records when the database learned that version. Suppose a project finished on Monday, but its completion report arrived on Wednesday. Asked today about the project’s status at the end of Monday, the agent can use the report to answer that it was complete. Asked what evidence a decision maker had on Tuesday about that same Monday, the agent must exclude the report. At that cutoff, completion was still unreported. Graphiti records why a relationship was revised. In Chalef’s example, contradictory evidence prompts Graphiti to add an invalidAt date and link the relationship to the source episodes responsible for its invalidation.
A correction must reach claims derived from the old wording, saved summaries, and searchable copies. It can also race a writer that read the old revision. Require the writer to identify its starting revision and enforce that precondition when committing the edit. HTTP’s If-Match is a concrete example: a server compares the supplied strong validator with the current representation before performing the mutation. An intervening change invalidates the writer’s basis for overwriting the record. Reread the new version and reconsider the evidence before proposing another edit. Software Engineering Fundamentals
Expiry, deletion, and reappearance
A low-ranked record can still be retrieved. An expired preference may remain available to a historical investigation. A read restriction can exclude a record before its stored copies have been removed. To specify forgetting, identify both the end condition and the action it triggers. A time to live sets a duration according to a particular clock. In LangGraph’s store interface, expiry support depends on the adapter, reads normally refresh the timer when TTL is enabled, and deletion is best effort. A frequently read obsolete preference could therefore survive indefinitely under a sliding timer. A policy tied to project closure needs an explicit closure event.
Follow deletion from source to assertion. In Chalef’s example, an intake episode is the sole source for a contact preference. Removing it also removes that preference. A reported allergy survives because two other episodes still support it. Now consider the wording of a shared summary. A surviving source may establish the allergy but say nothing about how to contact the person. Zep’s deletion documentation says that surviving shared-node names and summaries are not regenerated. To preserve only the supported information, rebuild the text from the remaining permitted sources and check each surviving detail.
Forgetting while indexing is underway
ExampleCleanup can finish while a job still holds old text.
Read the diagram as text
- Source session.
- Buffered text.
- Forgotten session ID.
- Tracked copy removed.
- Publication check.
- Source session → Buffered text: earlier read.
- Source session → Forgotten session ID: forget request.
- Forgotten session ID → Tracked copy removed: then clean up.
- Buffered text → Publication check: proposed write.
- Forgotten session ID → Publication check: exclusion.
OpenClaw removes the entire tracked entry if any merged source session is forgotten. The figure follows a job already holding that session’s text.
Report exclusion from future reads, completion of tracked cleanup, and residual copies separately. Check transcripts, exports, backups, and restore paths. If a copy must be retained for a restricted purpose, identify its owner and exclude it from ordinary reads and extraction. Reapply exclusions before a restored store can serve or publish. OpenAI’s Memory FAQ explains that retained older chats can generate memories again after memory is re-enabled. Deleting an application record also cannot retract an answer already delivered or change parameters trained on that information. Deletion propagation · Learned influence
Evidence of useful memory
Testing memory behavior
Build a test from an earlier interaction, the information permitted for retention, any intervening change, and a later task that does not give away the answer. Specify what should be available and what must be absent. Run independent cases against isolated histories. For a replay, freeze the information cutoff across originals, summaries, and indexes. Otherwise a summary can reveal a correction that arrived after the decision under test.
| Boundary | Discriminating test |
|---|---|
| Formation | Read the committed record beside the original. Check subject, qualifiers, and whether a detail excluded by policy was written anywhere. |
| Recall | Include a near-match for another user and an expired claim. Compare permitted IDs and versions with what the model received. |
| Correction | Pause writes. If the source is current but search returns old wording under the right note ID, compare the saved note with its indexed copy. Old wording in the note locates a problem before indexing. Current wording in the note directs attention to publication or serving. |
| Evidence use | Supply the correct permitted record directly. A wrong answer under direct delivery calls for inspecting interpretation, other inputs, and the task itself. |
| Forgetting | Pause a writer, forget its source, then resume it. Inspect the publication decision and stored derivatives. A clean answer alone cannot establish removal. |
In an oracle condition, the evaluator supplies the record known to be relevant to each task. Stefania Druga compared this with no recall, similarity search, and lookup from a ledger of prior decisions while keeping the model fixed. Even with the correct record in its input, the model sometimes ignored or misused it. If ordinary recall fails where direct delivery succeeds, inspect the saved record first. If it misstates the source, investigate the write. If it is accurate, check which revision reached the model and how the model acted on it. A trace that captures these successive states lets the investigator follow the error. Observability
Trace the suspected transition before changing the policy. LangChain’s Vivek Trivedy asks whether deterioration first appears after the first or second compaction. Compare the inputs on either side. Did the relevant detail disappear, or did it remain visible while the action changed? Once a fix is proposed, Ben Hylak’s replay approach suggests testing past behavior for side effects. A change that restores one missing detail should also be tested on cases where that detail ought to stay out.
LoCoMo labels the turns containing answer evidence, allowing retrieval to be assessed separately from the answer. LongMemEval tests cross-session reasoning, knowledge updates, and abstention when the history lacks an answer. Memora specifies both facts a response should use and obsolete facts it must exclude. These constructed histories answer behavioral questions. Claims about user isolation, deletion completion, and stale writers require tests of the application’s actual storage and access paths.
Measuring the value of recall
A memory system earns its upkeep through changes in later work. In Parth Asawa’s database task, an agent repeatedly answers questions about a database. Once it has discovered which tables join, it might answer the next question with fewer exploratory SQL queries. Score correctness alongside query count. Guessing a join saves a query too. Run the same agent on matched questions with experience retained and with experience reset. Report absolute success in each condition as well as the gain attributable to carrying state forward.
Compare a richer memory policy with a credible cheaper alternative, such as recent history or a small explicit profile. For a component comparison, hold model, tasks, tools, and budgets fixed and verify that recall actually ran. For a comparison of complete systems, report actual resource use and the stopping rule. In Sara Hooker’s AutoScientist example, stopping once a target had been passed affected the distribution of observed win rates. A policy allowed more attempts has changed the experiment even if the memory store is identical.
| Model | Without preference memory | With preference memory |
|---|---|---|
| Llama-3.3-70B-Instruct | 41.78% | 46.38% |
| Qwen-2.5-7B | 36.21% | 35.18% |
Llama’s users in the simulation also needed fewer preference reminders, down from 2.98 to 1.96 per session. Qwen’s success fell with memory enabled. In a separate study, 19 participants used assigned preferences over three sessions. Preference-related ratings improved, but participants described weaker transfer across domains, and some preferred to restate their choices. These conditions support a bounded comparison. They leave open how much time a person would save after months of maintaining and correcting a profile.
For a live comparison, assign users or workspaces to separate policies and isolate their histories. Switching policies per request would let one inherit the other’s writes. Separate tasks that depend on earlier experience from tasks where carryover should have no effect. Count correct continuations and repeated clarification alongside stale assumptions, inappropriate disclosures, and correction effort. Report read latency, write latency, and consolidation cost in their own units. Treat missing feedback as unknown: a user who stops correcting a system may have stopped relying on it. Evals and Benchmarks
Frequently asked questions
Does a longer context window give an agent cross-session memory?
A longer window increases the amount of information that fits in one model call. To use a decision from another session, the application must retain it and supply it again, either automatically or after a lookup. MemGPT’s paginated search is one way to bring just the relevant part of external history into the next call.
Does agent memory require a vector database?
No. A keyed record can supply a known preference, and a file can preserve a note for later reading. OpenClaw supports keyword search over retained files using SQLite. Vector search adds a way to find passages expressed in different words.
Can a deleted memory return?
Yes. A retained chat can supply the information again during extraction. A queued job can also recreate a deleted entry from text it read earlier.
Open questions
When does repeated behavior justify a standing preference? Several choices may share one temporary constraint. Progress would mean predicting which defaults transfer to new tasks while preserving exceptions and reducing the work users spend correcting them.
How much of a synthesis can survive selective forgetting? A remaining source can support the main claim while a single phrase exposes the removed source. A convincing result would regenerate the useful text from permitted evidence, verify the served revision, and withstand a delayed writer holding the old text.
Which policies save effort over months of real use? As people come to expect continuity, they change what they explain and which failures they report. A field comparison with isolated histories would need to measure correct continuation, inappropriate reuse, correction time, and operating cost together.
































































































































































