pi-diet-memory

One of my diet context engineering and workflow extensions. Explore the complete collection: https://www.npmjs.com/~probabilityengineer

Explicit lightweight persistent memory for Pi with no automatic context injection.

pi-diet-memory gives Pi agents a small, inspectable memory store for durable user preferences, project conventions, and lessons learned. Memories are saved only when explicitly requested or proposed for user approval. The extension avoids transcript scraping, vector databases, background consolidation, and automatic prompt injection, keeping prompt overhead low while still letting agents search memory when relevant.

How to use it

Repository visibility

list, search, and candidate review default to global memories plus memories for the current repository. The /memory, /memory-search, and /memory-review commands use the same visibility rules. Filtering happens before search and limits. Use scope: "global" for only global records, or scope: "project" for only current-project records. allProjects: true explicitly includes other projects (and still respects an explicit scope).

Project identity is the real, symlink-resolved Git root, including worktrees. Outside Git repositories it is the exact real working directory: child directories are separate identities, and moving or copying directories does not transfer their memories automatically. Project creation requires an existing directory.

Existing records are never rewritten during retrieval. Existing absolute legacy subdirectory paths are associated with their current Git root when resolvable; missing paths, relative paths, and records without project associations are hidden by default and remain inspectable with allProjects: true. A reused path or a changed Git boundary cannot be distinguished from its historical owner using the legacy path-only format; review such records explicitly after filesystem changes.

Changing global scope to project assigns the current identity; changing project scope to global removes the association. Other updates preserve the association. No memories are automatically injected into context.

Use pi-diet-memory for durable context that should survive sessions:

• user preferences
• workflow habits
• recurring corrections
• stable project conventions
• lessons learned

Do not use it for:

• secrets
• temporary task state
• raw transcripts
• one-off command output
• facts that are better stored in repo docs

Commands

Show saved memories:

/memory

Search approved memories:

/memory-search jj workflow

Save an explicit memory:

/memory-remember I prefer concise recommendations first, then caveats.

Review proposed memories:

/memory-review

Tool

One compact tool:

• memory — list, search, review, remember, propose, approve, update, and forget memories.

Save an explicit user-requested memory

Use remember only when the user asks to remember something.

{ "action": "remember", "text": "User prefers compact action-enum tools for small Pi extensions.", "kind": "preference", "tags": "pi,extensions,tools", "pinned": true }

Propose an inferred memory

Use propose when the agent notices a durable preference, but the user did not explicitly ask to save it.

{ "action": "propose", "text": "User prefers positive guidance phrased as desired behavior rather than negative bans.", "kind": "preference", "evidence": "User requested this style while discussing AGENTS.md guidance.", "tags": "guidance,style" }

Review pending candidates

{ "action": "review" }

Approve a candidate

{ "action": "approve", "id": "cand_..." }

Search approved memories

{ "action": "search", "query": "positive guidance" }

Update a memory

{ "action": "update", "id": "mem_...", "text": "User prefers concise positive guidance phrased as desired behavior rather than negative bans.", "pinned": true }

Forget a memory

{ "action": "forget", "id": "mem_..." }

Memory kinds

Use one of:

• preference — user style or workflow preference
• project — stable project convention
• lesson — correction or mistake to avoid
• reference — useful durable fact or link

Scopes

Default is global:

{ "scope": "global" }

Use project scope for repo-specific memory:

{ "scope": "project" }

Storage

JSON Lines under:

~/.pi/agent/pi-diet-memory/
  memories.jsonl
  candidates.jsonl

Memories are human-readable and easy to inspect or back up.

Safety

• Memory is context, not authority; current user instructions and repo evidence win.
• Do not store secrets.
• Prefer concise, stable preferences and lessons over raw transcript snippets.
• Use propose for inferred memories, review for pending candidates, and remember for explicit user requests.

Install

From npm:

pi install npm:pi-diet-memory

From GitHub:

pi install git:github.com/ProbabilityEngineer/pi-diet-memory

For project-local install, add -l:

pi install -l git:github.com/ProbabilityEngineer/pi-diet-memory

For local testing:

pi -e ./index.ts

Development

npm install
npm run lint