Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

prism-mcp

Source: github.com/dcostenco/prism-mcp Language: TypeScript | Status: v5.1.0, Very Early (solo author)

“Mind Palace” for AI agents — persistent session memory with behavioral learning, time travel, multi-agent sync, and a visual dashboard.

What It Does

prism-mcp gives agents persistent memory across conversations through three layers: an append-only session ledger (what happened), mutable handoff state with optimistic concurrency control (what’s current), and behavioral memory that learns from corrections (what to avoid). High- importance lessons can auto-graduate into .cursorrules / .clauderules.

Key Features

  • 30+ MCP tools across session, memory, search, and dashboard domains
  • Three-layer memory: session ledger, handoff state (OCC versioned), behavioral
  • Three-tier search: FTS5, sqlite-vec vectors, TurboQuant JS fallback
  • TurboQuant: pure-TS embedding compression (ICLR 2026) — 768-dim from 3,072 bytes to ~400 bytes (7x), >90% top-1 retrieval accuracy
  • Time travel via versioned handoff snapshots (memory_checkout)
  • Multi-agent hivemind with role isolation (dev/qa/pm)
  • Behavioral learning: corrections accumulate importance, auto-surface
  • Progressive context loading: quick/standard/deep tiers
  • Web dashboard at localhost:3000 (knowledge graph, timeline, health)
  • Morning briefings after 4+ hours of inactivity
  • SQLite (local) or Supabase (cloud) backends

Comparison to kb-mcp

Aspectprism-mcpkb-mcp
DomainAgent session memoryCurated document search
Content sourceAgent-generated session logsHuman-authored markdown
SearchFTS5 + sqlite-vec + TurboQuantBM25 (Tantivy)
Write modelAppend ledger + upsert handoffkb_write to writable collections
StorageSQLite or SupabaseTantivy index + disk reads
Unique featureBehavioral learning + time travelToken-efficient kb_context
Tool count30+10

Relationship: Different domains entirely. kb-mcp retrieves curated knowledge; prism-mcp persists agent session state. Complementary — an agent would use both simultaneously.

Patterns Worth Adopting

  • Progressive context loading — formalized quick/standard/deep tiers for kb_context could help agents pick the right depth.
  • Optimistic concurrency control — relevant if kb-mcp ever supports concurrent writers to the same collection.
  • Health check with auto-repair — extending kb_health to suggest or apply fixes, not just diagnose.

Source Metrics

ComponentLanguageFilesCode LinesCommentsTotal Lines
Core serverTypeScript6917,0126,41426,141
MigrationsSQL142,2276703,207
TestsPython152,1502682,754
DocsMarkdown91,0091,466
Total12128,3678,58740,977

Doc-to-code ratio: 0.05x. The codebase is large relative to documentation. Notable: single-author v1→v5 in 3 days with 96KB handler files suggests rapid feature accretion. The 30+ tool count is unusually high for an MCP server and may cause prompt bloat.