agent-memory: making a coding agent remember across sessions

An agent forgets what it learned when it closes a session. Which file it changed and why, which attempt failed: none of that is there next time. agent-memory is a runtime that closes that gap.

Memory tools have split into two lines. One is a retrieval engine built on embeddings and a knowledge graph; the other is a filesystem the agent reads directly. The first finds well but is hard to inspect, the second is legible but does not scale as a search surface.

This tool puts both in one store. The files stay as they are and an index sits on top, so the index searches and the files answer.

A memory is not a pasted block of transcript. It records what was done and why in one file, and links related files together.

Each memory is a Markdown file. Links inside a file carry relations, and files in the same folder are neighbours, so relations survive without a graph database underneath.

The index can be deleted at any time. Removing it and rebuilding loses no knowledge, and that is enforced by a test rather than promised in a document.

rm -rf .index/ && mem rebuild

It is not on PyPI yet, so you install from a checkout. It needs Python 3.12 or later, and installing builds the mem command, the MCP² server and the hook binaries.

uv sync --all-packages

After creating the store you attach it to the agent you use. Each host keeps its settings in a different place, so the command takes the host name.

A search returns a one-line abstract, the file path and a score, not the text itself. The agent looks at that list and opens only the files it needs, at the depth it needs.

Pasting text fills the context quickly and makes it hard to see what was read. Answering with paths leaves a record in the transcript of which files were opened and which were not.

It does not wait for the agent to decide to write something down. Distillation starts at the point where a conversation breaks, and it does not hold up the work in progress.

What is worth keeping usually surfaces when a task ends. The choice that was made and what failed get captured then.

The raw trace is copied before distillation. Something the distiller misses is therefore never lost by the system.

There is no model client inside the library. Judgment is borrowed from the host agent's CLI³, and it stays visible in the transcript.

A consolidation pass runs when there is no conversation. It merges what came in and forgets what has low value. It runs on its own cycle instead of waiting for a request.

What the unattended pass may and may not do is separated. It adds and updates on its own, but deletion only arrives as a proposal you confirm.

Forgetting is a feature. The same fact scattered across files duplicates search results, and an old decision can hide the current one.

Updating does not erase the earlier record. Asking as of a date answers with the memory as it stood then, so you can trace what changed when.

The developer measured a reduced LongMemEval-S with 120 episodes. The haystack was bounded to 12 sessions, the host was Claude Haiku 4.5 and a Sonnet 5 judge scored the answers.

These numbers do not measure writing quality or reasoning. They look only at how accurately facts from earlier sessions come back.

A memory written by Claude Code is found by Codex CLI and picked up again by Hermes. All nine ordered writer/reader pairs across the three hosts passed, with specifics intact.

Memory carries across hosts because the store is files. A store trapped inside one app's database cannot be read by another app.

Compared with starting without memory, correct answers rose from 2 of 36 to 13 of 36. Memory carries over even when the host changes.

The version is 0.1.0 and it is not on PyPI yet. Command names and settings can change, so follow the documentation that matches the version you installed.

The measurement bounded the haystack to 12 sessions. That makes it a study of write strategy rather than a comparison at corpus scale.

Switching agents means installing the hooks and settings again. The memory stays as Markdown, but which host writes and when depends on the settings.