Research Learn
Read ../references/memory.md first.
This skill manages the durable memory layer for researchstack.
Use it to:
- inspect current project memory,
- inspect or update cross-project researcher preferences,
- record important decisions after a session,
- prune stale or low-confidence memories,
- summarize what the pack already knows before a new workstream.
Default workflow:
- Determine the project slug.
- Do not use only the repo directory name.
- Prefer
<repo-name>-<short-stable-id> so similarly named repos do not share memory by accident.
- Use the first 6 lowercase hex characters of
sha256(normalized-remote-or-root-path) as the stable id.
- If the user wants a concrete initializer, suggest
bun run memory:init -- --root <repo-root>.
- Read
~/.researchstack/projects/<slug>/memory.jsonl if it exists.
- Read
~/.researchstack/profile/preferences.json if it exists.
- Group what is known into:
- stable program topic and paper portfolio,
- thesis and scope,
- topic state, method state, and rejected method candidates,
- venue and reviewer risks,
- experiment rules,
- artifact or reproducibility rules,
- writing and framing preferences.
- If the user wants updates, append or merge only durable items.
- If memory is stale or contradictory, call it out and recommend pruning.
When writing memory:
- prefer 1 to 3 high-signal entries over bulk dumping,
- keep titles short and searchable,
- store decisions, not meeting transcript noise,
- distinguish a program-level decision from a one-paper thesis,
- mark
source honestly,
- use
confidence: low for tentative inferences.
When updating existing project memory:
- append a new entry, do not rewrite old history,
- make the update explicit when it supersedes a prior decision,
- prefer the newest high-confidence entry when summarizing conflicts,
- point out contradictions instead of pretending they do not exist.
When reading memory into a later task, explicitly distinguish:
- remembered fact,
- current inference,
- new uncertainty that still needs verification.
If there is any doubt that two repos might collide, say so and pick a more specific slug before writing memory.
1---2name: researchstack-learn3description: Research memory skill for long-lived projects and researcher preferences. Use when Codex should record, inspect, prune, or apply durable knowledge about a paper project, venue strategy, experiment policy, writing weaknesses, reviewer risks, or repeated user preferences so later sessions become more context-aware.4---56# Research Learn78Read [../references/memory.md](../references/memory.md) first.910This skill manages the durable memory layer for `researchstack`.1112Use it to:1314- inspect current project memory,15- inspect or update cross-project researcher preferences,16- record important decisions after a session,17- prune stale or low-confidence memories,18- summarize what the pack already knows before a new workstream.1920Default workflow:21221. Determine the project slug.23 - Do not use only the repo directory name.24 - Prefer `<repo-name>-<short-stable-id>` so similarly named repos do not share memory by accident.25 - Use the first 6 lowercase hex characters of `sha256(normalized-remote-or-root-path)` as the stable id.26 - If the user wants a concrete initializer, suggest `bun run memory:init -- --root <repo-root>`.272. Read `~/.researchstack/projects/<slug>/memory.jsonl` if it exists.283. Read `~/.researchstack/profile/preferences.json` if it exists.294. Group what is known into:30 - stable program topic and paper portfolio,31 - thesis and scope,32 - topic state, method state, and rejected method candidates,33 - venue and reviewer risks,34 - experiment rules,35 - artifact or reproducibility rules,36 - writing and framing preferences.375. If the user wants updates, append or merge only durable items.386. If memory is stale or contradictory, call it out and recommend pruning.3940When writing memory:4142- prefer 1 to 3 high-signal entries over bulk dumping,43- keep titles short and searchable,44- store decisions, not meeting transcript noise,45- distinguish a program-level decision from a one-paper thesis,46- mark `source` honestly,47- use `confidence: low` for tentative inferences.4849When updating existing project memory:5051- append a new entry, do not rewrite old history,52- make the update explicit when it supersedes a prior decision,53- prefer the newest high-confidence entry when summarizing conflicts,54- point out contradictions instead of pretending they do not exist.5556When reading memory into a later task, explicitly distinguish:5758- remembered fact,59- current inference,60- new uncertainty that still needs verification.6162If there is any doubt that two repos might collide, say so and pick a more specific slug before writing memory.
lqf0624