Your AI forgets everything between sessions — who you are, what you decided, what failed, what works. Memory Crystal fixes that.
It captures each completed conversation once, distills durable knowledge only in bounded session-boundary or indexed nightly batches, and recalls the right context when an agent needs it. The turn write performs no inference. Capture stays automatic; recall defaults to deliberate tool calls so idle sessions cost nothing.
Point at https://api.memorycrystal.ai/mcp with a Bearer token
The universal installer supports cloud, local, and self-hosted backends. It can register MCP clients, configure OpenClaw, install Codex/Claude hook assets, and stage a Dockerized local Convex backend for offline or self-hosted work.
How it works
You send a message
│
▼
┌─────────────────────────────────────────┐
│ CONTEXT ENGINE │
│ │
│ Vector search across durable LTM │
│ Lexical fallback over recent messages │
│ Deterministic bounded ranking │
│ Scope filtering + context budgeting │
│ → Inject top memories into context │
└─────────────────────────────────────────┘
│
▼
AI responds with full context
│
▼
┌─────────────────────────────────────────┐
│ MEMORY EXTRACTION │
│ │
│ Completed turn → STM (idempotent) │
│ Bounded extraction → durable LTM │
│ One authoritative LTM vector index │
└─────────────────────────────────────────┘
Every response is informed by what came before. Every conversation feeds the next one.
Two memory layers
Layer
Stores
Retention
Short-term (STM)
Recent raw messages, verbatim
Rolling window by tier
Long-term (LTM)
Facts, decisions, lessons, people, rules
Permanent, vector-indexed
Sensory raw content
Raw sensory payloads behind memory records
Summarized, tombstoned, or protected by policy
STM gives recent continuity. LTM gives permanent knowledge. Sensory retention lets Memory Crystal keep durable recall value without keeping every raw payload forever.
Five memory stores
Store
Purpose
Example
sensory
Raw signals
"the user sounds frustrated about the deploy"
episodic
Events
"We shipped v2 on March 15"
semantic
Facts
"The API uses Convex for the backend"
procedural
How-to
"Deploy with npm run convex:deploy"
prospective
Plans
"Add billing webhooks next sprint"
Adaptive recall
Six modes, automatically selected:
Mode
Prioritizes
General
Broad recall across STM + LTM
Decision
Decisions, lessons, and rules before risky changes
Project
Goals, workflows, and implementation context
People
Ownership, collaborators, and relationships
Workflow
Procedures, rules, and how-to memory
Conversation
Recent session context and continuity
The Context Engine picks the right mode. You don't configure anything.
Knowledge bases
First-class immutable reference collections for docs, policies, runbooks, and imported source material. They sit alongside conversational memory so your agent can keep learned context and stable reference data separate.
Immutable — imported chunks stay stable, not rewritten by conversation
Scoped — tenant and scope filters keep KBs private to the right workspace
Bulk import — standard import or high-volume bulk-insert without blocking on embedding
Explicit retrieval — queried only when the agent selects the named KB
Bounded enrichment — embeddings are generated through guarded import/backfill jobs
Memory tools
Every tool works in any MCP host or automatically within OpenClaw hooks.
Tool
What it does
crystal_recall
Semantic search across all long-term memory
crystal_remember
Store a memory — decisions, facts, lessons
crystal_what_do_i_know
Everything known about a topic
crystal_why_did_we
Decision archaeology — why a past choice was made
crystal_preflight
Pre-flight check before risky actions
crystal_search_messages
Hybrid search over verbatim conversation history, with time-window (fromMs/toMs or ISO dates) and pagination
crystal_checkpoint
Snapshot memory state at a milestone
crystal_wake
Session startup — briefing and guardrails
crystal_trace
Trace a memory back to its source conversation
crystal_get_memory
Fetch the full text after crystal_recall flags a result truncated
crystal_recent
Recent messages for short-term context
crystal_edit
Update an existing memory
crystal_update
Update a memory in place
crystal_supersede
Replace an outdated memory while preserving history
crystal_forget
Archive or delete a memory
crystal_stats
Memory and usage statistics
crystal_health
Inspect bounded memory-hygiene work
crystal_create_knowledge_base
Create a named reference collection
crystal_list_knowledge_bases
List available knowledge bases
crystal_set_knowledge_base_access
Manage a KB's agent access
crystal_query_knowledge_base
Search a knowledge base
crystal_import_knowledge
Import reference chunks into a KB (upsert-by-dedupeKey replaces in place)
crystal_list_knowledge_base_memories
Enumerate a KB's chunks with cursor pagination + ids
crystal_empty_knowledge_base
Delete all chunks but keep the KB row/id/bindings
memory_search
Search LTM and return crystal paths
memory_get
Read a full memory by ID or path
Short-term message tools (crystal_search_messages, crystal_recent, recall message matches, and wake/session summaries derived from recent messages) redact likely secrets before returning tool output. A turn becomes searchable after capture finishes at turn completion; active same-turn tool calls should not expect the current prompt or response to appear yet.
HTTP API
All core operations available over authenticated HTTP:
POST /api/mcp/capture Create a memory
POST /api/mcp/recall Hybrid recall over all memory
POST /api/mcp/search-messages Search short-term history
GET /api/knowledge-bases List knowledge bases
POST /api/knowledge-bases Create a knowledge base
POST /api/knowledge-bases/:id/import Import chunks
POST /api/knowledge-bases/:id/bulk-insert High-volume migration
POST /api/knowledge-bases/:id/query Query a knowledge base
All endpoints require Authorization: Bearer <api-key>. Per-key rate limiting enforced.
Contributors can run an opt-in Dockerized Convex backend and dashboard locally, then seed fixture data without touching the managed production deployment:
npm run convex:local:up
npm run convex:local:seed
npm run convex:local:doctor
Provider keys are intentionally split: MEMORY_CRYSTAL_API_KEY is client bearer auth, GEMINI_API_KEY powers Gemini embeddings, and optional OPENROUTER_API_KEY powers durable memory extraction.
The local stack uses http://127.0.0.1:3210 for Convex RPC, http://127.0.0.1:3211 for HTTP actions, and http://127.0.0.1:6791 for the dashboard. See the Local-First Setup guide for end-user setup, local endpoints, backup and rollback behavior, and troubleshooting.
Self-hosted
Run everything on your own infrastructure:
git clone https://github.com/memorycrystal/memorycrystal.git
cd memorycrystal && npm install
# Deploy to your own self-hosted Convex backend
# Production Memory Crystal uses Railway self-hosted at https://convex.memorycrystal.ai
npm run convex:deploy
# Configure
echo 'CONVEX_URL=https://your-backend-url.example.com' > mcp-server/.env
echo 'GEMINI_API_KEY=your-key' >> mcp-server/.env
# Enable and verify
npm run crystal:enable
npm run crystal:doctor
For install-script OpenClaw installs, verify the loaded plugin directly:
openclaw plugins info crystal-memory
openclaw crystal_status
Multi-tenant isolation — owner checks on every retrieval, database-level separation
Per-peer / channel isolation — on multi-client accounts, memories and knowledge-base chunks are scoped by peer channel and fail closed for broad/absent scopes; recall, wake, and KB queries never surface one client's data in another's session
API keys — SHA-256 hashed at rest, plaintext never stored
