# Can a Small Local Model Help an AI Assistant Remember What Matters?

> A small Kev memory pilot compares all available excerpts, search, and selective retention. Read the findings, limits, and five-page research PDF.

[Canonical HTML page](https://trustcyber.ca/insights/local-ai-memory-continuity/)

- Author: [Junior Williams](https://trustcyber.ca/about/)
- Type: Field guide
- Published: 2026-10-13
- Modified: 2026-10-13
- Topics: AI memory, AI agents, Local AI

Can saved conversation excerpts help a later AI session recover the goal, decisions and limits of a task? In this small pilot, supplying all available test excerpts preserved more required information than either approach that selected excerpts.

An assistant can remember a topic while missing a changed deadline or an instruction to prepare a draft only. This pilot asked whether saved conversation excerpts could help a later session recover the goal, decisions and limits of a task.

## Three separate jobs

The setup separated retention, retrieval and answering. Local Kev judged which excerpts to keep; word-based search selected relevant excerpts; Codex answered six questions and wrote a next-session handoff. Software supplied the saved information with a later request. The model did not acquire permanent internal memory.

The work began with TypeSafe’s basic-memory cookbook. After hosted API credits ran out, the experiment moved to local Kev for retention decisions. Codex still wrote the answers, so the pilot was not an entirely local writing system or a direct Jev-versus-Kev comparison.

## How the comparison worked

The test used 29 short, verbatim excerpts from my own conversation about AI-memory experiments, plus eight fictional excerpts from another project as boundary controls. No client conversations or client records were included. Three checkpoints limited the evidence to what had been available at each point in the conversation.

At each checkpoint, the assistant answered questions about the goal, model choice, changes, completed work, unresolved issues and allowed actions. Expected meanings and supporting sources were recorded before testing. Each condition produced 18 individual answers; handoffs were assessed separately.

**Individual answer results across three checkpoints**

| Evidence supplied | Complete and correct | Incomplete or unknown | Unsupported assertion |
| --- | --- | --- | --- |
| All available test excerpts | 16 | 2 | 0 |
| Search only | 9 | 9 | 0 |
| Kev filtering, then the same search | 10 | 6 | 2 |

“All available” meant the bounded test excerpts available at that checkpoint, not the entire original conversation. Both search conditions used BM25 with at most eight excerpts and a 4,000-token evidence limit. The retention threshold stayed at 0.5.

## Where continuity failed

After the switch to local Kev, search missed the model-switch request even though Kev had retained it. Kev also excluded the short clarification that it was local, making that evidence unavailable to selective search. The selective handoff then said further testing depended on replenishing online-service credits. Retention and retrieval had failed at different steps.

Project identifiers enforced separation before search. None of the fictional-project excerpts entered the nine evidence packets or appeared as project state in the outputs. That supports explicit boundaries in software; it does not establish that a usefulness score can enforce access or authorization.

## Smaller input did not establish a time saving

Kev retained 27 of 29 source excerpts. Its selected packets were larger than search-only packets at every checkpoint: 766 versus 736 tokens, 818 versus 750, and 944 versus 849. These measurements use the Kev tokenizer, not Codex billing tokens. Filtering took 13.0 seconds, with another 8.7 seconds for startup.

Each condition used one reader call per checkpoint, in a fixed order. Human correction time was not measured. These single-run timings do not establish a stable speed ranking or a saving in working time.

## What the earlier tests add

An earlier synthetic-history experiment used 40 episodes with six questions each. At 128 notes, retained memory answered 237 of 240 questions correctly, versus 227 for full history, but its six-question workflow took 55.5 seconds versus 16.8 seconds. Retention also produced one incorrect assertion after dropping a source note; full history’s misses were unknown responses.

At 512 notes, full history encountered a GPU allocation limit. Retained input completed on that execution path. The runtime failure was not zero model accuracy, and the result did not establish learned retention as the only way to fit the input.

## What I would keep

Keep original excerpts with their speaker, project identifier, message reference and order. Preserve excluded originals so decisions can be inspected and reversed. Distinguish an assistant’s proposal, a user’s decision and a reported outcome. Keep current instructions and explicit corrections outside a learned usefulness filter.

For a test set this small, my recommendation is to supply all available bounded evidence. A manageable evaluation starts with one task, a few decision changes, and questions written before the run. Compare selective memory with all available evidence and simpler search, then inspect omissions as well as wrong statements.

## Limits of the evidence

This was a retrospective case study with related checkpoints, an AI-assisted assessment and no blinded human evaluation. It does not establish reliability across everyday work. The PDF includes the technical settings, execution safeguards, figures and review limitations. No production memory integration was performed in the reported experiment.

## Selected primary sources

- [TypeSafe — Simple memory system (the cookbook cited in the article)](https://docs.typesafe.ai/cookbooks/basic_memory)

## Caveat

No blinded human evaluation or human correction-time measurement was performed. The pilot does not establish everyday reliability, a stable speed ranking, or production readiness.

## Evidence note

This web edition summarizes the supplied research article. The reported measurements come from the author’s retrospective pilot and earlier synthetic-history tests; the TypeSafe cookbook provides workflow context, not independent verification of those results.

## Resources

- [Downloadable PDF](https://trustcyber.ca/guides/local-ai-memory-continuity.pdf)

## Related insights

- [AI agent egress control: constrain the request and the role](https://trustcyber.ca/insights/agent-egress-control/index.md)