Skip to content

memory.md

Latest commit

 

History

History
58 lines (40 loc) · 2.95 KB

memory.md

File metadata and controls

58 lines (40 loc) · 2.95 KB

Memory & Context Management

Plexir uses advanced techniques to manage long-running conversations, ensuring the model remains coherent even as the history grows.

Persistent Memory Bank (New in v1.7)

Plexir now includes a Long-Term Memory system powered by chromadb. This allows the agent to store and recall specific facts across different sessions.

Features

  • Semantic Storage: Memories are stored as embeddings, meaning the agent can find them even if the exact keywords don't match (e.g., searching for "database credentials" finds "db password").
  • Session Persistence: Memories persist even after you close Plexir. They are stored in ~/.plexir/memory.
  • Tools:
    • save_memory: The agent uses this to store explicit user facts (e.g., "The user prefers Python over C++").
    • search_memory: The agent uses this to recall information when needed.

Usage

You can prompt the agent to remember things directly:

"Remember that my API keys are stored in .env.local"

Or ask it to recall:

"Where did I say my keys were?"

Session-Scoped Scratchpad (New in v1.9)

For tasks that require multi-step planning and temporary note-taking, Plexir provides a scratchpad tool.

Unlike the global Memory Bank, the scratchpad is session-scoped. This means:

  • Isolation: Each session (started with a new run or loaded via /session load) has its own unique scratchpad file.
  • Plan Continuity: Plans made in the scratchpad are saved alongside your session. If you load a session tomorrow, the agent will see exactly where it left off in its plan.
  • No Global Noise: Content in one session's scratchpad won't confuse the agent in a different project or session.

The agent uses the scratchpad tool automatically for complex tasks, but you can also view it manually in ~/.plexir/sessions/<session_id>_scratchpad.md.

Rolling Summarization & Proactive Pruning

  1. How it works: Plexir identifies older, unpinned messages and uses the primary LLM to condense them into a concise "BACKGROUND SUMMARY."
  2. Context Preservation: This summary is injected at the start of the conversation, allowing the model to remember high-level decisions and context while clearing out detailed token-heavy noise.
  3. Automatic: This happens in the background without user intervention.

Message Pinning

You can manually protect critical messages from being summarized or pruned using Context Pinning.

Pin a message

Use the /session pin command followed by the message number (visible in the history or logs).

/session pin 5

Pinned messages will be highlighted and are exempt from the summarization process.

Unpin a message

/session unpin 5

Distillation (Failover)

During a Provider Failover (e.g., Gemini Primary hitting a quota), Plexir uses a "Distillation" process to transfer only the most essential recent context to the backup provider. This ensures a smooth transition with minimal latency and token waste.