rohitg00/agentmemory v0.7.4

v0.7.4 — reflect: auto-synthesize insights from memory clusters

on GitHub

What's New

Reflect — higher-order insight synthesis (#89)

The headline feature competitors like Hindsight charge for. memory_reflect traverses the knowledge graph, groups related memories by concept clusters, and synthesizes higher-order insights via LLM.

# Trigger via MCP
memory_reflect

# Or REST
POST /agentmemory/reflect

How it works:

1. Finds high-degree concept nodes in the knowledge graph
2. BFS clusters related concepts (depth=2)
3. Gathers semantic memories, lessons, and crystals for each cluster
4. LLM synthesizes cross-cutting insights (capped at 5 per cluster, 50 total)
5. Deduplicates via SHA-256 fingerprint — re-synthesis reinforces existing insights
6. Falls back to Jaccard document-similarity grouping when graph is empty

New type: Insight — confidence-scored, fingerprint-deduped, with daily decay sweep

New MCP tools: memory_reflect (in ESSENTIAL set), memory_insight_list
New REST: POST /reflect, GET /insights, POST /insights/search
Integrated: Runs as "reflect" tier in consolidation pipeline (after semantic, before procedural)

Also in this release

• #84 Fixed CrystallizableAction type error
• #85 npm package reduced from 40MB to 11.6MB
• #86 22 tests for lessons
• #87 11 tests for obsidian-export
• #88 Viewer: Actions tab (status filter, frontier highlights) + Crystals tab (card layout)
• #92 README function table: all 89 mem:: functions documented in categories

Stats

• 43 MCP tools (8 visible by default)
• 103 REST endpoints
• 627 tests

Install: npm install @agentmemory/agentmemory

Full Changelog

v0.7.3...v0.7.4