Skip to content
Skills Library
Personal productivity · Tool integration

memory

/memory

Use when the user asks about their projects, preferences, people, decisions, or past work; asks when something last happened; asks to remember or correct information; when a conversation establishes new information worth remembering; or when the user wants to see how their memories connect.

Licence
MIT
Last repository change
05.10.2026

Assessment and security

90/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content25 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
2
Scripts
0
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 05.10.2026, without references to retired models or APIs.

Licence

MIT, open: use and modify while retaining author attribution.

What it does and when to use it

Use when the user asks about their projects, preferences, people, decisions, or past work; asks when something last happened; asks to remember or correct information; when a conversation establishes new information worth remembering; or when the user wants to see how their memories connect.

Original author description

Use when the user asks about their projects, preferences, people, decisions, or past work; asks when something last happened; asks to remember or correct information; when a conversation establishes new information worth remembering; or when the user wants to see how their memories connect.

How it works

  1. Showing graph images
Original SKILL.md headings · 37 lines, 1,051 words
SKILL.md preview
---
name: memory
description: Use when the user asks about their projects, preferences, people, decisions, or past work; asks when something last happened; asks to remember or correct information; when a conversation establishes new information worth remembering; or when the user wants to see how their memories connect.
---

# Memory

Use the graph-memory MCP for persistent memory. Discover its tools before answering from past conversations or assuming that no memory is available. The connection chooses the memory graph; callers do not supply a namespace.

Consult memory before work that depends on the user's history, preferences, project conventions, or previous decisions, including ambiguous requests. Skip lookup for self-contained tasks that need none of that context. If repeated errors or unexpected behavior suggest a prior decision may matter, look up that context before repeating the same approach.

- Recall information with `memory_recall`, passing the subject as `entity` and the user's question as `question`. Start with the default compact response. Request more detail only when the returned facts leave a specific gap.
- Search remembered facts across entities with `memory_search`, passing the full question or a topic. Use this when the subject is unclear, the question spans several entities, or an entity lookup misses. This is keyword search; try fewer distinctive words or an alternative name if needed. An empty search does not establish that no relevant memory exists.
- Find names and identities with `memory_search_entities` when needed. Use a short name such as `Ketch` or `T3 Code`, not an entire question: all words must match the same entity's names or aliases. Search without a kind filter first; use returned entity keys for follow-up calls, and put the question in `memory_recall.question`.
- Use `memory_status` to check episode counts, processing, queued/failed work, and the latest saved or completed episode. Active leases do not prove worker liveness.
- Use `memory_latest` for when something last happened, `memory_evidence` for sources, and `memory_render` to visualize connections. Evidence accepts 1–10 fact IDs in one call: `detail:"index"` lists short claims and source IDs, compact returns bounded excerpts, and `detail:"full"` returns exact quotes and source context. Expand only relevant IDs; use full evidence before sourced writes or when an excerpt leaves a verification gap.
- Use `memory_ingest` to remember new information, `memory_retract` to withdraw an incorrect fact, and `memory_merge` for confirmed duplicate identities. Discover the input schema when using each tool.
- A conflict label can reflect an erroneous exclusive slot. When the original evidence or user establishes compatible alternatives, use `memory_allow_alternatives` to clear that grouping while preserving both facts. Do not retract a valid alternative merely to hide the conflict.

Keep lookup focused on the task. Follow returned entity keys or evidence references to resolve specific gaps; avoid exhaustive graph traversal or broad filesystem searches. Stop after targeted searches find no useful match.

Entity search requires a nonempty name; it cannot list everyone with an empty query. If a known entity's question returns no matches, try fewer distinctive terms or omit the question for bounded entity context; do not treat that fallback as an answer to the original question. Historical calls share a memory budget: on a busy response, retry sequentially after a short delay with the same cutoff. Do not silently replace a required historical cutoff with a live query.

Learn from the conversation as well as recalling it. When work establishes a durable new fact, preference, decision, or correction, use the memory tools to retain it within the user's authorized scope; an explicit "remember this" is not required. Keep useful knowledge rather than routine progress chatter. Preserve who said it and when, distinguish proposals from completed work, and reuse source/message identities on retries. Background workers already ingest saved conversations, so do not upload whole sessions again to make them learnable. An accepted write can still be queued and unavailable to recall.

Before writing a manual note, check whether recall already contains the original statement. Direct `memory_ingest` writes need a `sources` mapping from submitted message IDs to `source_message_id` values returned by full `memory_evidence`, using exact source excerpts. Do not turn your paraphrase into a fabricated user message or timestamp.

Interpret write receipts narrowly. `context_only_message_ids` identifies submitted messages that cannot produce searchable facts from this manual write; other sourced messages may still be processed. `available_for_recall:false` does not establish that the information is absent elsewhere or permanently lost. Say “Context accepted; searchable recall not confirmed.” Background ingestion is a separate path that may later capture the original conversation; neither promise that it will nor say the information cannot carry over. Check targeted recall before claiming a searchable fact was saved. Do not repeat unsourced writes or recommend editing instruction files solely because of this receipt; discuss a standing rule only when that is the user's goal.

Preserve the source roles: user statements and assistant reports are conversational claims; tool outputs validate claims rather than originate them. Memory-file summaries are derived material. An assistant report is not confirmed merely because it was saved or repeated. Apply corrections through the tools rather than rewriting source evidence or editing native memory files.

The graph represents recorded evidence, not a live check. Verify potentially stale external state when verification is practical; otherwise state that the answer is memory-derived and may be outdated. Use returned source references for claims about prior work. Ingestion time is not event time. To test what memory could have supplied before a conversation, set `known_at` to its start time; current recall can contain information learned during that same conversation. Pending or failed episodes and unseen sessions limit coverage; zero pending episodes alone does not establish that ingestion is complete. A search miss means no match was found, not that the person or relationship does not exist.

## Showing graph images

When the user requests a graph image, call `memory_render` with their requested scope or Cypher; omit Cypher for the whole graph. Keep the returned image and render metadata. An image visible in a tool response may not be visible to the user.

For Codex chat, save the returned image to a persistent local PNG outside tracked source files, then embed it in the final reply with `![Memory graph](/absolute/path/graph.png)` and provide a separate full-size file link. Use the actual saved path. Reuse the original render result instead of rendering again merely to deliver it.

Write image bytes through a file/binary API or stdin when available. If shell commands are the only option, decode and append bounded base64 chunks, splitting on multiples of four characters; never place the entire encoded image in a shell argument or Markdown data URL. Large renders exceed command-length limits. Report any truncation indicated by the render metadata.
Files in this package (2)
  • SKILL.md 7 KB
  • agents/openai.yaml 1 KB

Versions

RepositoryScoreLicence
This page · robertogongora/openclaw-graphiti-plugin90MIT
rjmurillo/ai-agents100MIT
rjmurillo/ai-agents100MIT
topoteretes/cognee-integrations100Apache-2.0
topoteretes/cognee-integrations100Apache-2.0
topoteretes/cognee-integrations100Apache-2.0
thebtf/engram100MIT
stxrmik/antigravity-plus100MIT
boshu2/agentops100Apache-2.0

Database snapshot: 13.11.2026 · package version ff9d16e322c5 (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

  • 90Memory recall and capture
    Use this skill when the user wants to recall something from past conversations, project decisions, or learned preferences across sessions — or when they want to explicitly save something to long-term memory. Trigger keywords include "remember", "recall", "what did we decide",…
  • 90remember
    Capture, correct, forget, and recall durable user preferences, decisions, and project context in an existing memory store or an explicitly chosen local folder. Use for explicit memory requests and relevant recall.
  • 90daftari-memory
    Use the user's Daftari vault as long-term memory. Use when the user refers to earlier decisions, preferences, or project context; asks you to remember, forget, or correct something; or when a conversation reaches a conclusion worth keeping for future chats.
  • 90remember
    Remembers durable things about Agora members and the group — preferences, habits, who does what, standing facts like allergies — and shows or forgets them on request. Use when a member says "remember that…", "note that I prefer…", "what do you remember about me?" or "forget…
  • 90memory-capture
    Save durable information to Mora only when the user explicitly asks to remember, save, or capture a fact, decision, insight, or task. Use for governed write-back, not for implicit extraction from ordinary conversation and never for deletion.
  • 90nemp-memory
    Agentic memory that evolves with your work. Use at session start or when the user wants to remember or update decisions and preferences, retrieve project context, or review saved memory.

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.