The memory API for every AI agent.
Evidence-backed semantic recall, immutable history, forced tenant isolation, and always-injected directives. Behind one API key.
Free tier · no credit card · REST + SDK + CLI + MCP
# store a memory, scoped to a user
curl https://api.memapi.in/v1/memories \
-H "Authorization: Bearer mem_sk_..." \
-d '{"userId":"u_123",
"text":"Prefers terse answers"}'
# => { "id":"fact_8fa2", "evidence":1 }- Free
- tier, no card
- 1024-d
- Voyage voyage-3.5 embeddings
- 4
- surfaces: REST · SDK · CLI · MCP
- 0
- cross-tenant leaks by design
Memories · user u_123
Prefers dark mode and terse answers
score 0.942 evidencepreferenceShips with Next.js 15 and Drizzle ORM
score 0.883 evidencecontextNever use em dashes in output
directivealways injected
Plugs into every harness
- Claude Code
- Cursor
- ChatGPT
- Kiro
- Codex
- VS Code
Not a vector store. A memory system.
Everything Mem0 does, plus the hard engineering it skips.
Always-injected directives
First-class rules assembled verbatim ahead of learned memory. Portable identity that follows the user across every tool, so you never re-teach a harness.
Evidence-backed recall
Every memory traces to its source. Explainable ranking you can inspect, not a black box.
Forced multi-tenant RLS
Row-level security on every tenant table. Memories and keys never leak across tenants.
Temporally immutable
Append-only history. Nothing is silently overwritten; you can always see what changed and when.
Async add + event polling, rich scoping
user / agent / run / app scoping ergonomics, async ingestion with an event stream, and an eval harness so retrieval quality is measured, not assumed.
Three calls. Any agent remembers.
The whole loop, from raw context to injected memory.
- 01
Add
Send raw text or events. MemAPI extracts durable facts, embeds them, and links every one back to its source evidence.
- 02
Bootstrap
Query at recall time. Directives assemble verbatim, memories rank by explainable relevance, all within a token budget.
- 03
Inject
Get a ready-to-use system prompt or structured context. Drop it into any agent, workflow, or app.
How MemAPI compares
We do not lead on the feature checklist. We lead on the posture: memory you can audit, correct, and prove.
| Capability | MemAPI | Maximem | Mem0 | Zep | Letta | Cognee |
|---|---|---|---|---|---|---|
| Byte-exact evidence span per factEvery fact resolves to the exact source quote, on screen, or is flagged unavailable, never reconstructed. | ||||||
| Correction retires the old factSuperseded facts leave the current set. Not accumulate-and-rerank. | ||||||
| Temporal validity on immutable revisions | ||||||
| Review-gated conflict resolutionA contested fact is a workflow item with typed actions, not a silent overwrite. | ||||||
| Verified deletion receiptsProvable erasure for GDPR / CCPA, not filtered-out-of-results. | ||||||
| Forced row-level tenant isolation | ||||||
| In-product retrieval debuggerPer-provider ranks and the citation trail, visible in the dashboard. | ||||||
| Open, run-it-yourself benchmarkMaximem shipped this first. We match it, with the evidence attached to every hit. | ||||||
| mem0-compatible drop-in surface |
Cannot a rival just copy this?
Any single row here is copyable in a sprint, and we expect it. What is expensive to copy is the architecture underneath. MemAPI stores byte-exact provenance, temporal validity, and review-gated correction at the schema level, not as a field bolted onto a recall-first store. Retrofitting that into a live vector database slows writes and complicates retrieval, which is why recall-first rivals resist it.
The durable moat is not the feature, it is the posture: owning auditable, correctable memory as a category, the switching cost of an agent whose correction history and evidence trail already live here, and shipping the honesty story (run the benchmark yourself, see why the agent believes it) faster than a rival defending inflated numbers.
See it live: why the agent believes a factGive your agents a memory today.
Free tier, no card. Full REST API, SDK, CLI, and MCP server.