pi-session-memory

SQLite search. Multi-agent analysis.

Find scattered history.
Analyze it together.

Search native Pi conversation history in one local SQLite index instead of relying on lossy summaries. Discussion activates historical Pi sessions as private participants. Save cited findings as reusable knowledge.

Install from npm Read the docs
pi install npm:pi-session-memory

Version 1.1.0 · Install from npm.

Multi-agent evidence routing

1 MAIN + 3 HISTORY
One current main agent routes recall to transcript evidence or up to three user-confirmed historical Pi agents. The main agent can independently use fetch_session to verify persisted context.
Discussion changes one recall result, not the main agent. Source transcripts remain unchanged.
Pi-native by design.One history formatSQLite retrieval · Discussion-ready sessions · Markdown output

Find scattered context

Ask to review earlier work. Pi searches native session history with SQLite-backed keyword and phrase matching. One session format keeps retrieval aligned with Discussion.

Explore retrieval

Multi-agent Discussion

Activate up to three user-confirmed native Pi sessions as historical agents. They answer privately and sequentially without changing source history.

Explore Discussion

Portable knowledge

Save cited findings, decisions, uncertainty, and actions as draft Open Knowledge Format (OKF) Markdown. Optionally use an Obsidian vault.

Explore saved knowledge

The 1.0 multi-agent architecture

One main agent.
Multiple historical perspectives.

Discussion routes one recall request through user-confirmed native Pi sessions. Pi-only history avoids external formats that Discussion cannot activate directly. The main agent keeps its tools and independently checks persisted evidence.

Read ADR 0001
1
current main agent
Up to 3
historical Pi agents
2
explicit evidence routes
Request-scoped routing Discussion changes the evidence response. It does not replace the main agent.
Current sessionMain agentowns tools, source checks, and final judgment
recall_memoryone explicit historical-review request
Route A · transcripts

Direct evidence review

Discussion off. Ranked transcript excerpts return to the main agent.

Multi-agent Discussion Route B · consultation

Ask historical Pi agents

A1History A2History A3History

User-confirmed agents answer privately and sequentially. They have no tools and never write to source sessions.

Compared by the main agentCited conclusionfindings · decisions · uncertainty · actions

Multiple perspectives for interpretation. Persisted transcripts for verification.

The review workflow

Evidence first. Conclusions second.

Historical review starts only when you ask for it. Ordinary questions do not trigger transcript recall.

Read the retrieval contract

  1. Check saved knowledge

    Inspect meeting descriptions. Reuse an existing OKF document when its description clearly matches.

  2. Retrieve historical evidence

    Search global history and entity-matched project sessions. Expand exact turns when more context is needed.

  3. Review or consult

    Compare retrieved evidence directly, or activate user-confirmed native Pi sessions for optional private Discussion.

  4. Save a cited conclusion

    Write draft minutes and portable evidence. Keep uncertainty explicit. Participant agreement is not verification.

Get started

Bring your history
into the next session.

You need Pi and native Pi session files. This shared format supports both transcript retrieval and Discussion activation.

Read the setup guide

  1. Install from npm

    Use the command above. Restart Pi or run /reload to load the extension.

  2. Choose your settings

    Ask Pi: Set up pi-session-memory. The guide explains one option at a time and saves only after your approval.

  3. Import your existing history

    Ask Pi: Import my Pi conversation history. You can also use the /pi-session-memory menu.

  4. Ask for an explicit review

    For example: Review our earlier work on offline eval and summarize the decisions and open questions.