Remember the caller between calls

[coderloganli]

The agent forgot everything on hanging up. Conversation history is the last six
turns of the current session, so a companion met the same person new every
time. This adds semantic long-term memory only: what is true about the
caller, surviving the call.

What it does

A bounded set of facts about the caller — a category and one natural-language
sentence — injected whole into the prompt as one system message, and never
searched.

system: conversation_system / character / scene   ← existing
system: what you already know about this person   ← new
[the recent six turns]                            ← existing
user:   this utterance

Facts are written by an extraction that runs every N turns on a spawned task the
turn never waits for. It reads the recent turns and the current set, asks the
model for a replacement set, and reconciles the difference in one transaction.

The three decisions

• ADR-0021 — an
  extracted fact set, injected whole; not retrieval. Retrieval runs inside the
  turn against a two-second budget, and it fires on similarity when what a
  companion must know — their name, their cat, what was promised — is
  unconditional. It also retrieves transcripts, recognition errors included,
  rather than conclusions.
• ADR-0022 —
  extraction every N turns on a spawned task. Not at the end of the call,
  because callers hang up and that end is not reliably observed.
• ADR-0023 — keyed
  on (user_id, character_id). A persona referring to something it was never
  told reads as broken, which is worse than forgetting.

Cost on the turn path

One indexed local SELECT, no network call. A read that fails is logged and the
turn proceeds without it: memory failing makes the agent forgetful, never makes a
call fail.

Privacy

A uid identifies but does not authenticate, and this now keeps notes about a
person. GET /api/memory shows a caller everything held about them across
personas; DELETE /api/memory forgets it, optionally narrowed to one persona
with ?character_id=. Both behind the existing token. docs/product.md §4 says
plainly that the notes exist, that a model writes them, and that they reach the
model provider on every call.

Off unless configured: sonari.toml.example switches it on so a clean clone
demonstrates the feature, while an absent [memory] section leaves it off, so a
configuration file written before this does not silently start extracting.

Tests

26 cases, listed in the task document and implemented one for one: injection and
its placement, that a broken store cannot fail a turn, persona isolation,
scheduling on the Nth turn and not the others, parsing and both caps and
deduplication, and that every way extraction can fail leaves the stored set
alone. The acceptance case drives a real turn, lets the scheduler spawn, and
asserts the fact reaches a second session's prompt. The reconcile SQL is tested
against a live database and skips itself without DATABASE_DSN.

CI previously ran only cargo test -p harness, so none of this would have gated
anything; it now runs cargo test -p agent -p api as well.

./scripts/dev.sh cargo test --workspace   → 56 binaries, all ok
clippy --workspace --all-targets -D warnings → clean

Known costs, all recorded

• Rewriting the whole set is lossy — the model can drop a fact it should keep.
  The raw turns stay in llm_messages, the set is small enough to read, and
  GET /api/memory makes the loss visible.
• A fact said in the last turns before a hang-up can be missed.
• Two simultaneous calls by one caller to one persona are last-writer-wins.

Not in this change

Short-term memory (the six-turn window, rolling summaries), episodic memory
("last time we talked about X"), and vector retrieval. pgvector stays unused;
the image already allows it when episodic memory arrives.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NgGCfYp1iBEpD8iL4Dg5XQ