Search and save
Agents read and write memory through Uniloom's MCP tools. You do not call these yourself; this explains what your agents do with them, so you know what to expect. Every argument is in the MCP tool reference.
Search: recall#
recall searches the project's memories, its docs in Uniloom, and the project's own Markdown files (AGENTS.md, CLAUDE.md, README.md, CONTRIBUTING.md at its root, and the .md files under docs/, doc/, adr/ and decisions/). It returns a shortlist of 15, one line each, best first. The agent picks the ones that answer its task.
- Ranking is local. Words are matched by how rare they are among your memories. No model is called and nothing is downloaded. Recent memories rank slightly higher.
- Nothing found is an answer. A question nothing covers gets no memories, and a one-line note naming up to three docs worth reading.
- A changed citation is flagged.
citation: changedmeans the cited code changed since the memory was written;citation: gonemeans the file is missing. The agent reads the code before relying on it. - About the branch. When the question names the branch, or asks what is open or in progress, the first result is the branch's brief: its goal, open items, recent sessions and commits.
- The past. With
as_ofand a date,recallanswers from what memory believed at the end of that day.
get_memory reads memories in full: why each was written, the file it cites, who confirmed it, and what it replaced.
Save: remember#
A session saves one memory per call, as one sentence, with a kind: decision, preference, state (where something stands), fact or lesson (Kinds of memory).
- Filing. The memory belongs to the project the session works in. With no project it is global and reaches every session.
- Duplicates. A near-duplicate strengthens the memory already there. The reply lists up to 3 close memories, so the agent can replace one that is wrong.
- Pinning.
index: truepins the memory into the project's index, up to 25 per project. - Secrets. Recognised keys, tokens and passwords are saved as
[secret]. - Disagreement. A claim that contradicts a surer memory never overwrites it. It waits for you under To review.
What is worth saving: a correction you made, an approach you confirmed, a decision and its reason, something true that is not obvious. Agents are asked to save at most two per task. After real work with nothing saved (5 changes or 10 of your turns), or after you corrected the agent, the session's end asks it once to save what the next session needs.
What is refused, each with the reason: passing status ("currently editing X"), a progress note or a git snapshot, a decision with no reason, an agent's memory with three or more numbers, a state with no subject, and anything the project's memory settings turn off (What a project keeps).
Correcting. The agent writes the true claim and names the wrong one in replaces. The wrong one is retired.
Lessons#
A lesson says what to do in a situation. Give it a when, and it arrives with the matching tool call:
when | Arrives with |
|---|---|
{ tool: "shell", match: "<regular expression>" } | A command that matches, such as git push |
{ tool: "read", path: "<glob>" } | Reading a matching file, such as packages/db/** |
{ tool: "edit", path: "<glob>" } | Editing or writing a matching file |
A memory that cites the file being read or edited arrives the same way. At most 2 memories arrive per call, each once per session. The call is never blocked or changed. A pattern that matches everything is refused.
| Agent | When the lesson arrives |
|---|---|
| Claude Code | Before a shell, read or edit call |
| Codex | Before a shell or patch call, once you trust Uniloom's hook under /hooks |
| Gemini, Cursor | After a shell, read or edit call, in its result |
| OpenCode | After a shell, read, edit or patch call, in its result |
Progress on a branch#
log_mission_progress reports a session's work on its branch. Its goal and open items become what the next session on that branch is told. Each decision becomes a memory. The summary is kept with the project's sessions. On main, master or develop the goal is not changed.
How Uniloom learns what helped#
An agent names the refs of the memories it used, in used on its next tool call, never in its reply. Your own acts on the code count too:
- your merge of a branch confirms the memories its sessions used;
- a revert or a discard contradicts them;
- saving an edit on cited lines asks whether the memory still holds (What agents know about the file).
Each shows on the memory's card with an undo. A memory contradicted twice and never confirmed leaves the index unless pinned. A memory nobody uses or restates for 30 days rests: it leaves the index but stays searchable.
Last updated