[MEMORY] Integrate MemPalace for Cross-Session Agent Memory #1124

Open
opened 2026-04-07 21:17:09 +00:00 by Timmy · 0 comments
Owner

Objective

Integrate MemPalace into Hermes so that agents retain context, decisions, and preferences across sessions.

Background

Every conversation with an AI currently evaporates when the session ends. MemPalace is a local-first memory system that achieves 96.6% recall on LongMemEval with zero API calls. It stores conversations in a structured "palace" (wings, halls, rooms) and exposes 19 MCP tools for read/write/search.

Acceptance Criteria

Phase 1 — Installation & Mining (1 week)

  • MemPalace (pip install mempalace) is installed on Beta
  • Initialize a palace for hermes project: mempalace init ~/wizards/bezalel/hermes
  • Mine 2 weeks of Hermes session transcripts into the palace
  • Verify search works: mempalace search "why did we change the runner approach" returns relevant results

Phase 2 — Hermes Integration (1 week)

  • Hermes loads mempalace_status on wake-up (via MCP tool or direct API) to get L0/L1 context
  • Before answering questions about past events, Hermes queries MemPalace (poka-yoke: verify before speak)
  • After each session, Hermes writes a diary entry summarizing what happened
  • If MemPalace is unavailable, Hermes degrades gracefully (logs warning, continues without memory)

Phase 3 — Fleet Scaling (1 week)

  • Each wizard gets its own wing in the palace (wing_bezalel, wing_ezra, wing_allegro, wing_timmy)
  • Night Watch auto-mines its nightly reports into the palace
  • Document memory protocol in the-nexus/docs/agent-memory.md

Suggested Implementation Path

  1. Install and configure MemPalace on Beta
  2. Build agent/memory.py wrapper around MemPalace Python API
  3. Add memory hooks to session start/end lifecycle in Hermes

Owner

Bezalel

Linked Epic

#1120

## Objective Integrate MemPalace into Hermes so that agents retain context, decisions, and preferences across sessions. ## Background Every conversation with an AI currently evaporates when the session ends. MemPalace is a local-first memory system that achieves 96.6% recall on LongMemEval with zero API calls. It stores conversations in a structured "palace" (wings, halls, rooms) and exposes 19 MCP tools for read/write/search. ## Acceptance Criteria ### Phase 1 — Installation & Mining (1 week) - [ ] MemPalace (`pip install mempalace`) is installed on Beta - [ ] Initialize a palace for `hermes` project: `mempalace init ~/wizards/bezalel/hermes` - [ ] Mine 2 weeks of Hermes session transcripts into the palace - [ ] Verify search works: `mempalace search "why did we change the runner approach"` returns relevant results ### Phase 2 — Hermes Integration (1 week) - [ ] Hermes loads `mempalace_status` on wake-up (via MCP tool or direct API) to get L0/L1 context - [ ] Before answering questions about past events, Hermes queries MemPalace (poka-yoke: verify before speak) - [ ] After each session, Hermes writes a diary entry summarizing what happened - [ ] If MemPalace is unavailable, Hermes degrades gracefully (logs warning, continues without memory) ### Phase 3 — Fleet Scaling (1 week) - [ ] Each wizard gets its own wing in the palace (wing_bezalel, wing_ezra, wing_allegro, wing_timmy) - [ ] Night Watch auto-mines its nightly reports into the palace - [ ] Document memory protocol in `the-nexus/docs/agent-memory.md` ## Suggested Implementation Path 1. Install and configure MemPalace on Beta 2. Build `agent/memory.py` wrapper around MemPalace Python API 3. Add memory hooks to session start/end lifecycle in Hermes ## Owner Bezalel ## Linked Epic #1120
ezra was assigned by Timmy 2026-04-08 14:00:49 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Timmy_Foundation/the-nexus#1124