agent-memory

An agent that closes its session forgets everything it learned in it, and the two usual fixes both cost something: a vector-and-graph retrieval stack that hands back an opaque chunk you cannot inspect or migrate off, or a flat markdown filesystem that stays legible but stops ranking the moment the tree grows past a listing. agent-memory, built by tigerless-labs and released in September 2026, tries to be both at once. Plain markdown files under one store are the single source of truth; a local, rebuildable SQLite index sits beside them purely as a ranked-retrieval cache. Claude Code, Codex CLI and anything else that can shell out share the same store, so a decision recorded in one host is findable from another.

Retrieval answers with an L0 list, one line per hit with a path, abstract and score, and the agent opens only as deep as the task needs: outline, then full file, then the cited raw session transcript. Three independent read paths back each other up: deterministic injection of a root MEMORY.md at session start, BM25 full-text search over an FTS5 index, and, if both miss, the plain directory an agent can ls and grep by hand. Writes happen at conversation boundaries rather than depending on the agent remembering to record something, and an optional sleep-time pass consolidates and prunes on its own schedule, but only ever proposes deletions for a human to confirm, never deletes unattended.

The natural fit is a team standardizing memory across more than one coding-agent CLI that wants the store itself to stay auditable: every memory is a markdown file with frontmatter, diffable in git, greppable without any tool, and provably reconstructable, since deleting the index and rebuilding it is enforced by a test to lose nothing. No LLM client lives inside the library itself, so there is no API key to configure and no separate billing surface; the reasoning is borrowed from whichever host CLI is already running, which also means every write it makes is visible in that host's own transcript rather than happening invisibly.

The project is three weeks old at the time of writing, with no packaged release yet, so installing means cloning and running uv sync from source. Its own benchmark section is unusually careful about what it does not prove: the 120-episode LongMemEval-S study it reports is explicitly flagged as not comparable to published LongMemEval numbers, because the haystack is bounded to keep it a write-strategy study rather than a corpus-size one. Worth separately noting: the tigerless-labs organization has shipped a new open-source tool roughly every two to three weeks through 2026, and several of them reached thousands of stars within weeks of creation, a launch cadence and star velocity worth watching, even though the commit history here includes real externally authored pull requests fixing actual bugs.

agent-memory is worth a look if you run Claude Code and Codex side by side and want one inspectable, portable memory store behind both, rather than a black-box retrieval service. The engineering is more careful than the marketing suggests, the write-up owns its benchmark's limits rather than hiding them, but three weeks of history, no packaged release, and a fast-moving parent organization mean it is still a bet on a young project rather than a settled choice. Treat the retrieval-accuracy numbers as an interesting internal study, not an independent result, until the project or a third party reproduces them.