titen.dev

AI agent memory
that can explain itself.

Agent memory that runs with no API key, no LLM, and no embedding provider. Every memory keeps its source, who may read it, and the evidence that contradicts it.

Cross over Read the docs

$curl -fsSL https://titen.dev/install.sh | bash

$titen mcp

No account, no API key, no server, no model. It opens ~/.titen/memory.db and speaks MCP over stdio.

Niteni

Javanese: niteni is to notice, and to keep what you noticed. An isekai premise is a memory problem: you arrive with nothing, and everything depends on what you can recall.

titen.dev

The other side

Three record types.

Titen never flattens what it stores: a conclusion and the evidence it came from are never the same record.

Observation

Raw evidence with a content hash and a time it occurred.

obs_ed91a2d0… · sha256 bba5cf65…

Claim

A conclusion that cites its evidence. Disputes stay disputed: both sides are kept, never averaged.

claim_f3963d7b… · verified · team

Context

Exactly what an agent was handed, why each record was selected, and what the budget cut.

ctx_ad0c00be… · 104/1,200 tok

The climb · level 1 → 6

Six ways to remember. Only two can explain why.

This is Titen's product vocabulary, not an industry standard. Pick a level to see what an agent keeps, and what it loses.

Level 01 · session

Remembering by re-reading

recall("pricing decision")

The mechanism

The window moves forward. Old turns fall away, and reading the file again is the only route back.

Nothing outlives the session.

Level 02 · retrieval

Remembering by similarity

nearest(query, 4)

The mechanism

The query becomes a vector. Nearby chunks return because they look alike, not because they remain true.

Match .79 was superseded in March.

Level 03 · structure

Remembering with types

memory.graph("pricing")

The mechanism

Memories gain types and links. The graph explains what relates, but not which source deserves trust now.

Structure is useful. Authority is still missing.

Level 04 · maintenance

Memory that tidies itself

consolidate({ subject })

The mechanism

Duplicates merge and low-value memories decay. Maintenance lowers noise without proving the result.

A cleaner memory can still be wrong.

Level 05 · kernel

Memory that can be checked

compile({ task, budget: 1200 })

The mechanism

Scope, trust, time, provenance, and feedback decide what enters a bounded context pack.

Every selected claim resolves to evidence.

Level 06 · fabric

Memory a team can share

handoff("agent-3")

The mechanism

Agents share one evidence kernel through scoped identity, leases, checkpoints, handoff, and audit.

Collaboration never widens private memory.

Level 5

Memory kernel

Everything one agent needs to build context: evidence, claims, temporal state, a context compiler, and feedback.

Evidence with provenance and content hashesClaims that cite, supersede, and expireContext compiled under an explicit budget

Level 6

Collaboration layer

Many agents share knowledge, keep private perspectives, and hand off safely. Leases keep them from clashing.

Identity and scope: org → workspace → project → runCheckpoints, task leases, handoff, and outcomeFull audit: who knew, wrote, used, and shared

Two homes, one core

Runs on the edge,
or entirely on your box.

The same kernel, the same API, the same export format.

Apache-2.0v0.10.0 · stable channelone package0 depsMCP at /mcp

quickstart

# self-host · Bun + SQLite$ titen bootstrap --org 'My Org'$ titen serve

# edge · Cloudflare Workers + D1$ wrangler d1 create titen$ pnpm deploy:worker

# then, from any agentconst ctx = await titen.compile({ task })→ 1 claim · 1 evidence · 104/1200 tok

The first useful slice

Five operations prove
the whole loop.

Observe, derive, compile, act, feed back. The five routes below are the whole loop.

Memory kernel · RESTMCP ready

observePOST /v1/observationsappend-only evidence

derivePOST /v1/consolidationsrules first, model only if it must

compilePOST /v1/context/compilescope, then rank into a budget

feedbackPOST /v1/context/:id/feedbackoutcomes tune future recall

traceGET /v1/claims/:id/evidenceevery claim resolves to sources

Vectors are an index, never the source of truth. Retrieved memory is reference data, never an instruction.

capabilitycloudflarevps

HTTPWorker fetchBun.serve

Canonical storeD1bun:sqlite

Lexical retrievalFTS5FTS5

Semantic vectorsVectorize · test stack onlysqlite-vec

Background repairCron Triggerin-process timer

Today, honestly

Semantic vectors are live on Bun/SQLite through sqlite-vec. On Workers, retrieval is lexical FTS5 until Vectorize lands, and claim extraction is deterministic on both. No model in the loop yet. FTS5 is the default, not a fallback: on the pooled 19,829-session LongMemEval-S condition (n=500) it outscored Titen's own FTS+vector arm, 0.246 against 0.212 — vectors are recommended for scoped, per-instance stores.

Today, honestly

Semantic vectors are live on Bun/SQLite through sqlite-vec. On Workers, retrieval is lexical FTS5 until Vectorize lands, and claim extraction is deterministic on both. FTS5 is the default, not a fallback: on the pooled 19,829-session LongMemEval-S condition (n=500) it outscored Titen's own FTS+vector arm, 0.246 against 0.212.

Measured, honestly

Pooled 19,829-session single-subject LongMemEval-S store, n=500: Titen FTS-only recall@1 0.246, significantly above every measured competitor (Mem0 OSS infer=False 0.182, p = 0.0007). Two pre-registered falsifiers fired against Titen in the same run: the 0.70–0.85 prediction was wrong, and compile p95 864.9 ms crossed the 250 ms line. The subject-scoped condition holds 0.880 at p95 138 ms — scoping is worth +63.4 points. Read the full benchmark.

Memory Atlas

See why an agent
remembered something.

A read-only view. Trace evidence, spot conflicts and stale claims, and see what was withheld from you as a count, never as content.

Open the Atlas

claim_f3963d7b…verifieddepth 2 · 28/40 nodes

The p95 latency budget for checkout stays at 400 ms.

supportsload-test-0714.json · p95 383 ms

qualifiesApplies to checkout only

disputedagent-11 says 250 ms · conf 0.44

withheld1 source exists but is not visible to you

Agent memory guides

Start with the problem you need to solve.

Three practical guides: the memory model, safe sharing between agents, and the MCP connection.

Cross over Docs

AI agent memory
that can explain itself.

Three record types.

Observation

Claim

Context

Six ways to remember. Only two can explain why.

Remembering by re-reading

Remembering by similarity

Remembering with types

Memory that tidies itself

Memory that can be checked

Memory a team can share

Memory kernel

Collaboration layer

Runs on the edge,
or entirely on your box.

Five operations prove
the whole loop.

See why an agent
remembered something.

Start with the problem you need to solve.

AI agent memory

Shared memory for AI agents

MCP memory server

AI agent memory that can explain itself.

Three record types.

Observation

Claim

Context

Six ways to remember. Only two can explain why.

Memory kernel

Collaboration layer

Runs on the edge, or entirely on your box.

Five operations prove the whole loop.

See why an agent remembered something.

Start with the problem you need to solve.

AI agent memory

Shared memory for AI agents

MCP memory server

AI agent memory
that can explain itself.

Runs on the edge,
or entirely on your box.

Five operations prove
the whole loop.

See why an agent
remembered something.