Your LLMs forget everything between sessions. Remember Ninja gives them structured, versioned, searchable memory — via REST, WebSocket, or MCP. No LLM on the backend. Pure heuristics. Sub-2ms reads.
LLMs lose everything when the session ends. Chat summaries are lossy, unstructured, and unsearchable. Your agents make the same mistakes twice because they can't remember what was decided — or why.
Most memory stores what was decided. Not why. Without reasoning, memory is a flat lookup table — not experience. The second layer is what turns memory into judgment.
Redis or JSON files give you no history, no provenance, no search. When a fact changes, the old one vanishes. You can't ask "what did we used to believe?"
Claude, Cursor, ChatGPT, OpenClaw, custom agents — each needs memory, each talks a different protocol. You're rebuilding the same persistence layer over and over.
Every stored fact is a versioned assertion with a canonical keypath, status lifecycle, provenance, and optional vector embedding. Memory isn't gospel — it's experience. Assertions are the accumulated output of previous thinking. Was the last decision good? Did the context change? Your agent can compare, supersede, and evolve what it knows.
No LLM on the backend. Retrieval is pure heuristics — deterministic lookups, full-text search, vector similarity, reciprocal rank fusion. No inference in the hot path. Sub-2ms reads. No GPU costs. Predictable latency at any scale.
{
"keypath": "decisions.auth.provider",
"value": "Auth0",
"context": "Evaluated Auth0, Clerk, and Okta. Chose Auth0 for SSO support."
}Not just the decision — the reasoning behind it. Canonical keypath, timestamp, actor trail, and the why.
# Exact keypath — O(1), sub-2ms
GET /v1/assertions/key/decisions.auth.provider
# Hybrid search — lexical + semantic
POST /v1/assertions/search
{ "query": "why did we choose this auth provider" }Keypath lookups for known facts. Hybrid search when the agent needs to explore past reasoning.
// PUT /v1/assertions/:id/supersede
{
"value": "Clerk",
"context": "Auth0 raised prices 3x. Migrating to Clerk."
}Old fact → superseded. New fact → active. Full chain preserved. That's experience, not just storage.
Each codebase gets isolated memory. Frontend decisions don't bleed into backend.
Each product in your SaaS suite has its own memory space.
Every end user gets scoped memory. Isolated, searchable, GDPR-erasable in one call.
Share memory across products in a group. Internal tooling patterns, company-wide decisions.
Scopes compose naturally. User within project within product group. Memory resolves at the right level.
Assertions capture what was decided and why. Agents use memory as accumulated judgment — compare outcomes, detect shifts, evolve understanding.
Pure heuristics. Deterministic lookups, FTS, vector similarity, rank fusion. Sub-2ms reads. No GPU costs. Predictable latency at any scale.
Agents stay connected. Persistent, bidirectional memory channel. Store and retrieve mid-conversation. Sub-5ms reads.
Active, superseded, retracted. Full history. Nothing silently overwritten. "I changed my mind" is a first-class operation.
Four modes: exact keypath, lexical FTS, semantic vector, hybrid with reciprocal rank fusion. Conflicts surfaced automatically.
12 tools via Model Context Protocol. Claude Code, Cursor, and any MCP client discover memory automatically. Zero glue code.
Hard-delete per scope or keypath pattern. Compliance receipts with zero PII. Built in, not bolted on.
Append-only event log. Every mutation tracked: who, what, when, why. Cursor-based pagination. Enterprise-grade provenance.
The memory backend is pure heuristics. No LLM. But the people managing it need intelligence. The admin panel uses AI for analytics and steering — pattern detection, drift monitoring, anomaly surfacing — while the memory layer stays fast and deterministic.
Drift detection across scopes. Memory health monitoring. Usage patterns. Anomaly surfacing when assertion patterns change unexpectedly.
Inject baseline knowledge before agents start. Global standards, per-project context, per-account preferences. Template scopes for new projects.
Browse, search, and edit assertions across all scopes. View supersession chains. Retract or correct facts when agents get it wrong.
API key rotation and revocation. Webhook configuration. GDPR erasure with dry-run preview. Full event log. Per-service usage metrics.
| Operation | Cloud API | CLI (local) |
|---|---|---|
| Keypath read | < 2ms | < 1ms |
| Write | < 5ms | < 10ms |
| Lexical search | 10–20ms | < 15ms |
| Hybrid search | 60–120ms | 60–220ms |
| MCP tool call | — | 1–10ms |
| Auth overhead | < 1ms cached | — |
The Remember Ninja CLI is free and open source. Install it, run it on your own machine with a local SQLite database. No server, no sign-up, no network calls. Full assertion model, full search, full MCP server — completely offline.
Use it standalone with OpenClaw or any MCP-capable tool for free local memory. Or connect the same CLI to the remember.ninja cloud service when you need shared, multi-user, cross-device memory.
npm install -g @remember-ninja/cli
remember init# Store a decision with reasoning
remember put "decisions.stack.framework" "Fastify" \
--context "Benchmarked against Express and Hono"
# Search your memory
remember search "framework decision"
# View decision history
remember history "decisions.stack.framework"{
"mcpServers": {
"remember-ninja": {
"command": "remember",
"args": ["mcp", "start"]
}
}
}# Same CLI, now syncing to remember.ninja
remember config set api.endpoint "https://api.remember.ninja"
remember config set api.key "your-api-key"Any HTTP client. No SDK required. TypeScript SDK available for convenience.
npm install @remember-ninja/clientPersistent bidirectional memory channel. Sub-5ms reads. The recommended path for agents that need constant memory access.
{
"transport": "websocket",
"endpoint": "wss://api.remember.ninja/v1/ws"
}12 tools auto-discovered. Zero glue code. Your LLM assistant gets persistent memory in one config block.
{
"mcpServers": {
"remember-ninja-cloud": {
"url": "https://api.remember.ninja/v1/mcp/sse"
}
}
}Plug Remember Ninja into OpenClaw as the shared memory layer across sessions and agents. Every agent in the swarm reads and writes to the same scoped memory. WebSocket keeps all agents in sync. Works with the free local CLI or the cloud service.
Claude Code and Cursor forget your project decisions every session. With Remember Ninja, every session recalls not just "we use Postgres" but why — and whether that decision worked out.
Each turn, your agent stores what it learned and retrieves what it knows. New information supersedes old, with reason captured. Over time: real experience, at heuristic speed.
Every agent in the swarm shares consistent context. Canonical keypaths as source of truth. When one agent updates a fact, others see the change with reasoning attached.
Thousands of users, each with scoped memory. Preferences, history, context — searchable, versioned, GDPR-erasable per user in one API call.
What did your AI believe at any point in time? Append-only event log and supersession history. Complete, immutable record with full provenance.
Inject company policies, architectural standards, and compliance rules before agents start. Starting points they can evolve — not locked-in rules.
You're wiring LLMs into products. You need memory that's structured, searchable, and doesn't lose history. Remember Ninja gives your agents a real memory layer — not a chat log dump.
Claude Code and Cursor are powerful, but they forget your project decisions every session. The free CLI gives them persistent, project-scoped memory that survives context window limits.
Your team builds with LLMs. You need auditability, compliance, and a memory layer that doesn't become another integration to maintain. One service, three transports, full provenance. Admin panel with AI analytics.
Full on-premise or private cloud installation. Same API, same performance, same MCP integration — running inside your security perimeter.
Contact salesFull on-site or private cloud installation
Same REST + WebSocket + MCP API surface
Your database, your network, your compliance
Dedicated onboarding and deployment support
SSO/SCIM integration for your identity provider
SLA, support agreements, audit trail included
REST: any HTTP client, minutes. MCP: add a config block, zero code. CLI: npm install -g, one command.
Yes. The CLI is open source and runs entirely on your local machine with SQLite. No account, no server, no limits. Use it standalone with OpenClaw or any MCP client. Optionally connect it to the cloud when you need shared memory.
No. Lexical search works out of the box. Semantic search is optional — plug in OpenAI, Voyage, or a local model if you want vector retrieval.
Export your full state as JSON or YAML at any time. The CLI stores everything in a local SQLite file you own.
MIT licensed. Standard protocols (HTTP, WebSocket, MCP). Your data is structured JSON with keypaths — portable by design.
Yes. Any framework that supports MCP or HTTP can use Remember Ninja as shared memory. Works locally (free CLI) or via the cloud. WebSocket for real-time sync.
Yes. Pre-seed assertions at any scope — global standards, per-project context, per-account preferences. Agents can supersede pre-seeded facts as they learn.
Hard-delete per scope or keypath pattern. Compliance receipts generated automatically. No PII in receipts.
Free forever. Local SQLite. Full assertion model, search, and MCP server. No account needed.
Get the CLIFree tier available. Usage-based pricing. Shared memory, WebSocket, admin panel, GDPR tools.
Start freeFull installation on your infrastructure. On-premise or private cloud. Dedicated support and SLA.
Contact salesStop losing decisions and their reasoning between sessions. Give your AI a memory layer that's fast, structured, and gets smarter over time — without an LLM in the loop.