{"id":23727,"date":"2026-09-04T17:43:19","date_gmt":"2026-09-04T17:43:19","guid":{"rendered":"https:\/\/scannn.com\/give-your-coding-agents-a-memory-you-own\/"},"modified":"2026-09-04T17:43:19","modified_gmt":"2026-09-04T17:43:19","slug":"give-your-coding-agents-a-memory-you-own","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/give-your-coding-agents-a-memory-you-own\/","title":{"rendered":"Give Your Coding Agents a Memory You Own"},"content":{"rendered":"\n

\nI work across several machines, and I switch coding agents depending on the task.
\nEvery one of them meets my projects as a stranger. The reasoning from \u201clast Tuesday\u201d
\ndisappears when the session ends. Each new agent, on each new host, starts from zero.<\/p>\n

Earlier this year, Software Forgets: Agent Traces Are the
\nMemory<\/em><\/a> made the case
\nthat coding agents already produce the record we keep losing. As they search a
\ncodebase, try approaches, hit errors, read documentation, and change direction, they
\nleave behind a dense account of not just what<\/em> changed, but why<\/em>.<\/p>\n

While the diagnosis is correct, traces are only potential memory. The session logs of
\nan agent are still just an archive. You cannot grep<\/code> your way to \u201cwhy did we move
\noff the streaming parser?\u201d<\/em> across ten thousand turns. For an agent to use those traces
\nwhile it works, they need indexing, retrieval, ranking, and exact provenance.<\/p>\n

That is what funes<\/a> provides. It is a durable
\nmemory layer for your agents (Claude Code, Codex, pi, and Hermes). It is built from the
\nsessions already on your machine. It works locally and becomes part of your agent’s
\nnormal workflow with one command. When you want it to, it can also travel to a Hugging
\nFace dataset you own, private by default.<\/p>\n

\n\t<\/p>\n

\t<\/a>
\n\t
\n\t\tAdd memory to the agent you already use
\n\t<\/span>
\n<\/h2>\n

funes is a single binary. Its default inference backend has no ML runtime dependency,
\nand embedding and reranking happen on your machine<\/em>. Install it:<\/p>\n

curl -fsSL https:\/\/huggingface.co\/buckets\/huggingface\/funes\/resolve\/install.sh | sh\n<\/code><\/pre>\n
Then add it to an agent:<\/p>\n
funes add claude    \n<\/code><\/pre>\n
That one add<\/code> command builds the first index, gives the agent recall<\/code> and get<\/code>
\ntools, and installs the automation that indexes each completed turn. Indexing is
\nincremental, with new runs adding new turns rather than embedding the whole history
\n again. The older and deeper content can backfill in bounded steps.<\/p>\n
From there, you just work. When a task touches a past decision, rationale, or finding,
\nthe agent can reach for recall<\/code> itself. You do not need to remember the old session or
\npaste its context into the new one.<\/p>\n
<\/a><\/p>\n
\n
With funes added, recall happens inside the conversation. The agent reaches for its
\nmemory on its own and names the session behind its answer.<\/p>\n<\/blockquote>\n
recall<\/code> returns the original text, not a summary, and shows exactly where it came from
\n(the agent, timestamp, session, and turn). Each result includes a get<\/code> command that
\nopens the full turn and its surrounding context.<\/p>\n
Underneath, one deterministic pipeline parses every supported trace into the same
\nturn-and-block shape, chunks it, embeds it with a pinned local model, and writes it to
\na local Lance<\/strong> dataset. A query combines vector and BM25 search, fuses their
\nrankings, reranks the candidates with a cross-encoder, reweights them by recency, and
\nattaches neighboring chunks.<\/p>\n
That design gives funes three important properties:<\/p>\n