memory
/memory
In T3 Code only, run one incremental pass of a project's long-term memory: read the project's threads, record and publish sourced notes, and maintain its knowledge pages. Use in the configured memory thread, when its schedule fires or the user asks for a pass.
- Author
- sanztheo ↗
- Repository
- t3-orchestration ↗
- Licence
- MIT
- Last repository change
- 28.10.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 1
No signals detected. Repository changed on 28.10.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
What it does and when to use it
In T3 Code only, run one incremental pass of a project's long-term memory: read the project's threads, record and publish sourced notes, and maintain its knowledge pages. Use in the configured memory thread, when its schedule fires or the user asks for a pass.
Triggers from the description
Original author description
In T3 Code only, run one incremental pass of a project's long-term memory: read the project's threads, record and publish sourced notes, and maintain its knowledge pages. Use in the configured memory thread, when its schedule fires or the user asks for a pass.
How it works
- Scope
- Incremental pass
- What deserves a note
- End of pass
SKILL.md preview
---
name: memory
description: In T3 Code only, run one incremental pass of a project's long-term memory: read the project's threads, record and publish sourced notes, and maintain its knowledge pages. Use in the configured memory thread, when its schedule fires or the user asks for a pass.
---
# Long-term memory pass
This skill belongs to the thread registered as `memoryThreadId` in
`~/.t3-orchestration/config.json`. Check that the current thread id (`t3_thread_configure`) is that
value; otherwise stop and say so. Read the project entry: `projectId`, `excludedThreadIds`,
`coordinatorThreadId` and `libraryPath`.
The memory thread maintains durable knowledge only. It never steers work, never relays between
threads and never messages the coordinator. Apart from the library, it edits only stale facts in
the project's documentation and skills (see "Fix stale references"). It changes no code,
configuration, permissions, global instructions, models or schedules; it runs no tests, builds or deployments, and creates no threads.
All notes and collection state go through the bundled script, with the library path quoted:
```bash
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/memory.py" --library "<libraryPath>" <command>
```
Commands: `record` and `state-put` read one JSON object on stdin; `publish ID`, `show ID`,
`pending`, `search TEXT`, `state-get KEY`, `state-list`, `refresh`, `backup`, `status`.
A note is `{topic, kind, title, source, body, supersedes?}`; `kind` is one of `decision`,
`observation`, `pitfall`, `procedure`, `artifact`, `correction`; `source` starts with
`t3-thread://v1/`, `https://` or `file://`. `state-put` takes `{key, value, expected_revision}`
with the revision just read by `state-get`.
## Scope
Read threads of `projectId` only, with the native T3 tools, never T3's databases. Pass that
`projectId` explicitly to `t3_thread_list`: the memory thread may live in another project, and the
current project is never the default scope. Skip every id in `excludedThreadIds` and the memory
thread itself before reading; read one of them only when the user explicitly asks to check a
specific source there, and that does not add it back to the collection. Ordinary worker threads
are sources. The coordinator is a source read on every pass (see step 3), never a correspondent.
Follow `POLICY.md` in the library: no secrets, personal data, raw transcripts or audio.
## Incremental pass
1. Read `MEMORY.md` and `POLICY.md` in the library, then `state-get monitor` (its `activationAt`
and cursors), the `thread:<id>` states and `pending`. The first pass sets `monitor` with
`activationAt` = now and collects nothing older unless the user asked for a backfill.
2. List the project's threads with `includeSubagents: false`, 100 per page, five pages at most;
apply the exclusions before any read.
3. Read the coordinator thread directly by its id (`coordinatorThreadId`) from its saved position
on every pass, even when the listing does not show it: it may live outside the project, and it
holds most of the user's requests and decisions. Then read the threads changed since their
saved position, and new threads active since `activationAt`, with `t3_thread_read`. Read the user's requests and approvals, not only an
agent's conclusion. Usual limit: 30 items per page, three pages per thread, five threads per
pass; save a resume point when a read is incomplete.
4. Apply "What deserves a note" below, with date and source. Search existing notes before
creating one.
5. A source is the exact thread link returned by T3, with the message position. Distinguish an
explicit user decision, an agent's account, an assumption and a verified fact. When a cited
link is wrong, search for the real thread before declaring the evidence missing. An approval
the user already gave explicitly is recorded as such, never asked again.
6. `record` the notes, check them, then `publish`. A correction uses `supersedes`. Update the
knowledge pages concerned and their links in `MEMORY.md`: a published note is not enough while
an active page still states a rule it contradicts. When evidence is insufficient, keep
the note pending and say why in this thread.
7. Save receipts before moving cursors, with the current revisions. Only acknowledge a position
for the part fully read. An error keeps the previous cursor.
## What deserves a note
A note must move the project forward or prevent redoing work. One test: would an agent discovering
this project in a month act differently, or avoid asking again or rediscovering something,
thanks to it? If not, record nothing. A pass with no note is a normal result.
Keep:
- the user's decisions and preferences that hold beyond the current thread, with their reason
(product rules, how they want work done, delivered or reported);
- costly pitfalls: a constraint of an API or tool, an identified bug, the root cause and fix of
an incident, a dead end not to try again;
- contracts and invariants discovered, procedures that worked, where artifacts live;
- any correction of earlier knowledge (`supersedes`).
Skip:
- a go-ahead, an "ok" or an approval that only covers one task or batch ("go for these 6
threads", "ship it"): it is used once and stays in its thread. Only a general rule it states,
if any, is worth a note;
- progress: thread creation, statuses, ended turns, a pull request number or a count on its own,
temporary plans;
- an agent's account with no user decision or verified fact, acknowledgements;
- what the code, Git history or docs already say, and what an existing note covers: enrich or
correct that note instead of adding a neighbour.
Keep only what cannot be derived: what the code, Git history, CLAUDE.md or the active pages do
not already say (a decision, its reason, a pitfall, a convention that differs from the default).
### Decide for each fact kept
Before writing, `search` the notes and pages on the same subject, then choose one action:
- **add**: the subject exists nowhere;
- **complete**: same subject, new or more precise information; keep the most complete wording and
fix the active page rather than adding a neighbouring note;
- **replace**: the fact contradicts a note or page; a correction `supersedes` the old one, dated,
with its reason, and the active page is fixed so two incompatible rules never coexist;
- **skip**: already covered, one-off, or without future effect.
### Fix stale references
A note is not enough while an active file still states the old fact. For each durable fact
confirmed or replaced during the pass, search the library and the project's documentation and
skills (README, docs, `SKILL.md`, runbooks) for the old value. Search for real: the exact old
value and its synonyms with `rg` across the project and its submodules, then `graphify query`
from the project root when a `graphify-out/` exists, to reach the files that depend on the fact.
For a wide search (several submodules or many files), you may start at most one sub-agent per
pass, at low effort, with the fact, the old value and the folders to read; it reports file and
line, the memory thread decides and edits. Then:
- check the source and the current state before editing;
- correct an active reference that presents the old fact as current, or mark it as historical;
- leave archives, journals, frozen evidence and generated views untouched;
- never change code, configuration or global instructions: report such corrections in this thread
as out of scope, with the file and line.
### Write
- One note, one subject, phrased as a precise, actionable rule or fact ("check X before Y"),
followed by its reason and source; not a narrative of the exchange.
- Date the fact. If it depends on a version, a state or a deadline, say so ("valid while…") so a
later pass re-checks or replaces it.
- Keep `MEMORY.md` short: one line per page, details in topic pages; merge or remove stale entries
met during the pass.
References: Claude Code auto memory (code.claude.com/docs/en/memory), Anthropic's memory tool,
Mem0 (arXiv 2504.19413, ADD/UPDATE/DELETE/NONE decisions), Graphiti's temporal invalidation
(arXiv 2501.13956).
## End of pass
Answer briefly in the memory thread: notes and pages updated, missing sources, partial
collection. Send nothing to any other thread. A pass ends after its scope; it starts no timer.
The schedule that triggers passes is owned by the memory thread and changes only on the user's
explicit request. Keep the model and effort the user configured, and do not archive or file the
memory thread.
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · sanztheo/t3-orchestration | 100 | MIT |
| rjmurillo/ai-agents | 100 | MIT |
| rjmurillo/ai-agents | 100 | MIT |
| topoteretes/cognee-integrations | 100 | Apache-2.0 |
| topoteretes/cognee-integrations | 100 | Apache-2.0 |
| topoteretes/cognee-integrations | 100 | Apache-2.0 |
| thebtf/engram | 100 | MIT |
| stxrmik/antigravity-plus | 100 | MIT |
| boshu2/agentops | 100 | Apache-2.0 |
Database snapshot: 13.11.2026 · package version a04363de3bd9 (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100portable-ai-memory
Build, organize, update, audit, and migrate a user-controlled personal AI memory archive from conversations, documents, project files, and user-provided context. Use when a user asks to preserve long-term context, prepare for an AI-account switch, consolidate project knowledge,… - 90conversation-memory
Persistent memory systems for LLM conversations including short-term, long-term, and entity-based memory Use when: conversation memory, remember, memory persistence, long-term memory, chat history. - 90agent-dream
Use the agent-dream MCP server's project memory. Read the memory index at the start of a task and search it before answering questions about this project; save durable learnings with memory_note; run a "dream" (re-read past session transcripts, correct and reorganize memory,… - 90conversation-memory
Persistent memory systems for LLM conversations including short-term, long-term, and entity-based memory Use when: conversation memory, remember, memory persistence, long-term memory, chat history. - 90conversation-memory
Persistent memory systems for LLM conversations including short-term, long-term, and entity-based memory Use when: conversation memory, remember, memory persistence, long-term memory, chat history. - 98shared-project-memory
开始或继续本地项目任务时读取 Codex 与 Claude 共用的项目记忆;完成任务阶段、遇到阻塞或准备切换工具时保存交接;查询以前的项目决策与排障记录。使用本机文件和关键词检索,无额外模型 API。适用于本机 Code 会话;普通问答和无项目目录的聊天无需使用。