Your AI Agent Keeps Making the Same Mistakes - Sparsh Jain, CentralAgent AI

Loading slides…

Read the talk

How an AI Agent Can Learn From the Engineer Who Corrected It

Sparsh Jain, Head of AI at CentralAgent AI, explains how his team turned corrections into reusable engineering judgment, preserved the history behind revised decisions, and cut repeat context gathering from 20 minutes to two.

From a talk by Sparsh Jain

At a glance

Ideas worth remembering

  • Preserve a correction at the level the evidence supports. Keep the detailed case, record a strategy with its conditions, or extract a skill the agent can reuse.

  • Use active records for routine retrieval and make superseded judgments available for deeper investigation. Give obsolete material a path to be forgotten.

  • Evaluate retrieval in the circumstances of the request. Two accurate records can imply very different actions, as the San Francisco and New York addresses illustrate.

  • Observe the memory lifecycle before automating edits. Extract lessons between sessions, test proposed changes independently, and check whether the next teammate benefits.

Ten engineers gain experience. Their shared agent repeats the mistake.

Imagine ten engineers working with the same AI agent for two years. Between them, the engineers have accumulated twenty years of experience. Each has seen decisions go wrong, intervened, and learned something to carry into the next job. What has the agent learned? Sparsh Jain poses the question because his team at CentralAgent AI had worked with the same agent for two years and still saw familiar mistakes recur.

A senior executive at a hedge fund described the failure to Jain in terms of a handoff. Someone who recognized a mistake could correct the agent and get useful work out of the current session. Start another session, or put the agent to work with another team, and it could make the same mistake again. The organization had people capable of correcting individual errors. What it could scarcely afford was to have someone recognize and correct every recurrence.

Jain’s team initially gave the agent all the context it could gather and relied on the agent to work out which facts mattered. With a small body of material, adding context helped. At enterprise scale, collecting and processing it took more effort, time, and tokens, and accuracy eventually fell. That led the team back to what had changed for the engineer during a correction. The engineer had exercised judgment, seen a problem with the original approach, and revised it. The team wanted that revision to survive the session.

A blue inequality sign separates “More Context” from “Better Intelligence” in a two-line heading. Three bullets below name tokens, latency, and performance.Open full source frame
More Context ≠ Better Intelligence
0:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:00 · section reference included

Working memory, durable facts, and three forms of experience

Jain reports a tenfold performance gain after rebuilding the system. To explain the design, he first separates short-term and long-term memory. Short-term memory is the working memory of the active session, the material the agent has in hand as it works. Long-term memory has two parts. Fact-based memory holds information the agent can draw on later. Experiential memory holds lessons abstracted from what happened when the agent and its users actually did the work.

The team gave those lessons three forms. The case base preserves a detailed situation together with the detailed learning from it. This is useful when a later reader needs the circumstances to understand why a decision made sense. The strategy base captures a conditional judgment. Given a situation of this kind, what approach worked for the team, and what should the agent try when it encounters those conditions again?

The skill base holds something the team can generalize well enough for the agent to use repeatedly. The distinction is the reach of the lesson. A correction can be worth keeping in full even if the team has only seen that particular case. A strategy identifies the conditions under which an approach has proved useful. A skill carries the reusable part of the work forward once the team can separate it from the particulars of the original experience.

Three illustrated panels pair cases with a stack of cards, strategies with a route on a folded map, and skills with a multitool. Each panel asks a question.Open full source frame
The Missing Layer: Experiential Memory
2:39
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:39 · section reference included

A revised judgment needs a history

Once the team could preserve experience, it faced a problem that a fresh memory store conceals. The data changes, the systems the agent works with change, and the surrounding environment changes. An approach that works today may be a poor choice a month or a year later. After a series of experiments, Jain’s team organized the memory lifecycle around three operations he calls formation, updation, and retrieval.

Formation captures a fact or an experience in a memory record. Updation keeps the collection current as further evidence comes in. The system can add a node for something newly encountered, revise an older node, or deactivate one that has ceased to be useful as guidance. Retrieval determines which of those records will inform the task in front of the agent. One decision about updating turned out to matter particularly for retrieval. What should an investigator see after the team has changed its mind?

Double-headed arrows connect three illustrations. Media icons feed into a stack under “Formation,” arrows circle layers under “Updation,” and a robot points to an image under “Retrieval.”Open full source frame
Formation, Updation, and Retrieval

Keeping only the final judgment creates a way to repeat an old error. An agent investigating the issue later can reason its way back to the earlier conclusion, unaware that the team already reached and revised it. Jain therefore distinguishes two depths of lookup. Surface-level research uses active nodes. Deeper research can consult active and inactive nodes together. Inactive nodes preserve the sequence of earlier judgments, so an investigator can trace the route to the current one without putting superseded advice into every routine session.

4:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:05 · section reference included

Why the memory lifecycle includes forgetting

Jain illustrates the cost of indiscriminate recall with the case of Solomon Shereshevsky, a man known for his extraordinary memory. In Jain’s telling, a choice most people could make in a second might bring a hundred recollections to mind, each seeming as relevant as the next. Shereshevsky could recall the details. The difficulty was deciding which one mattered for the choice in front of him.

An agent can accumulate the same burden. Every obsolete fact or incidental observation it brings forward is another item to weigh on the next task. Deactivation serves a particular purpose when an earlier conclusion explains a later decision. Forgetting serves a different purpose when material has ceased to be true or relevant and recalling it contributes little. The system needs a way to make both decisions, or the cost of working through accumulated context returns.

6:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:13 · section reference included

A model capable of checking proposed changes

At this stage the team had agents that could use experience and a memory system whose contents could evolve. Maintaining it still took manual work. Jain asks what had changed in the preceding six months to make the next step possible. His answer concerns a particular use of stronger models. A model could now serve as a verifier, taking on review work that had previously required a human to inspect each item.

That gave the team a way to build review into the system itself. Extracting a plausible lesson from a session and deciding whether it belongs in live memory could become separate operations. The verifier would have a place in the recurring evaluation process Jain describes later.

7:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:42 · section reference included

A correct address sends the agent to the wrong coast

For Jain, retrieval sets a ceiling on the performance of the whole system. He asks the audience to suppose he has homes in San Francisco and New York. Leaving his San Francisco office, he asks his agent to pull up directions home. Midway through the hypothetical journey, he realizes the route is taking him somewhere unexpected. The agent has picked his New York home.

Both stored addresses are correct. The word “home” fits either one. The decisive information is where he is leaving from and which home he means on this occasion. An agent can generate directions competently from the record it receives and still fail the request. The same dependency applies to a carefully maintained strategy. Forming the right lesson and keeping it current only help when retrieval supplies it in the circumstances where it applies.

Jain gives a measured comparison for work the team repeated. Gathering context for the agent had taken about 20 minutes in many cases. When the same agent had gathered that context before and knew what to gather again, the step took about two minutes. That is the scope of the tenfold timing improvement. Jain also reports lower token cost and less effort, with better performance on long-horizon tasks because the agent could reuse decisions the team had already thought through.

8:43
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:43 · section reference included

What happens during a dreaming cycle

Jain calls the recurring maintenance passes dreaming cycles. His analogy is a night’s sleep. Between periods of active work, less important material can be cleared away, items deserving attention can move forward, and the remainder can be reorganized. An agent that keeps acquiring context needs time designated for those operations or its priorities will increasingly reflect what happened to accumulate.

Clockwise arrows connect Observe at the top, Reflect at the lower right and Evolve at the lower left, then return to Observe.Open full source frame
Dreaming Cycles: How Agents Learn While You Sleep

He draws three operating rules from the work. First, disclose memory progressively, starting with what the present task calls for and making more available when deeper inquiry warrants it. Second, keep improving the system as its environment changes. Third, put the maintenance procedure into a loop. A prompt can guide an individual session. A loop specifies when the system will revisit what it learned and how a proposed change reaches the next session.

The build order matters. The team first implemented formation, updation, and retrieval, then added observability around those operations. On top of that it built the loop and evaluation layer. After a session or at the end of a day, the process extracts lessons, proposes improvements, tests them independently, and then applies the changes in the live system. Jain leaves the detailed evaluation design and the security guardrails for a separate discussion.

10:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:28 · section reference included

The next teammate is where learning pays off

Return to the ten engineers. In one calendar year, each person directly experiences one year of work. A shared agent can participate in work with all ten. What changes the handoff is that a judgment acquired with one engineer can now be available when the agent encounters a similar problem with another. Over time, its range of experience can extend across domains that no individual teammate regularly works in.

Jain says he has already seen teams of people and agents working together this way. He describes some as “10X or 100X” teams. The figure he puts a clock to is the time spent gathering familiar context, which fell from twenty minutes to two. His forecast for the next six to nine months goes further: he expects more autonomous AI companies in which agents collaborate and carry what they learn from one interaction into the next. For the hedge fund executive, the immediate test is simpler. When a different teammate starts a fresh session, can the agent put the earlier correction to work?

12:54
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:54 · section reference included

Transcript download
Full transcript

Loading transcript…