Titentiten.devInstall
The practical guide

AI agent memory, from context window to shared evidence.

AI agent memory is the system that lets an agent retain, retrieve, and use information beyond one model call. Useful memory also records where a claim came from, when it was valid, and whether it helped.

Install TitenView source
6 levels
from session context to collaboration
3 records
observations, claims, compiled context
2 runtimes
Cloudflare/D1 and Bun/SQLite
01

What is AI agent memory?

A context window carries the current turn. Agent memory survives it. The memory system decides what to store, how to retrieve it, and what belongs in the next bounded prompt.

That makes memory more than chat history. An agent may need durable evidence, a current claim derived from that evidence, a checkpoint for unfinished work, and feedback about whether the last recall was useful.

Titen keeps observations, claims, and compiled context separate. The distinction preserves the raw source while allowing conclusions to change over time.

  • Observations preserve what a user, tool, or system produced.
  • Claims cite observations and can be disputed, superseded, revoked, or expired.
  • Compiled context selects authorized claims under an explicit token budget.
Read the Titen model →
02

Six levels of agent memory

Titen uses six product levels to describe what a memory system can do. These levels are a design vocabulary, not an industry standard.

  • Level 1 keeps session context and raw files.
  • Level 2 retrieves semantically similar chunks from an external store.
  • Level 3 adds typed memories and relationships.
  • Level 4 extracts, consolidates, and forgets automatically.
  • Level 5 compiles evidence-grounded temporal context and learns from feedback.
  • Level 6 adds identity, visibility, coordination, governance, and optional federation.
Inspect all six levels →
03

Why a vector database is not the whole memory system

Vector search answers which chunks look similar to a query. It does not decide whether a chunk is still true, who may see it, which source supports it, or what must be cut when the prompt budget is full.

A production memory path needs a canonical store behind the index. Retrieval should consider scope, trust, time, provenance, conflicts, feedback, and budget before an item reaches the model.

Titen treats vectors as an index. On Bun/SQLite, sqlite-vec provides semantic retrieval. On Cloudflare Workers, retrieval remains lexical FTS5 until Vectorize is shipped.

See the current vector boundary →
04

A production checklist for agent memory

Start with the failure modes. Ask what happens when two sources disagree, a fact expires, a private record is queried by another agent, or a recall fills the context window.

The smallest useful production loop is observe, derive claims, compile context, act, and record feedback. Shared systems then add separate credentials, visibility, leases, checkpoints, handoffs, and audit.

  • Keep raw evidence append-only and addressable.
  • Make temporal state and supersession explicit.
  • Filter unauthorized records before ranking.
  • Return provenance with every selected claim.
  • Record outcomes so future compilation can improve.
Review the product invariants →