3tears-agent-intention

3tears Agent Intention

Standing-wants corpus for LLM agents. An intention is agent-authored deliberation output -- a want the agent is holding onto ("ask Pace about the migration", "check whether the wake threads are firing") -- with its own status lifecycle, salience decay, and restraint brakes so the agent surfaces a want at most occasionally rather than every turn.

Part of the 3tears framework.

Installation

pip install 3tears-agent-intention

Why a separate package (not a memory of kind=intention)

Memory is a hard-delete-only user-fact taxonomy with immutable scope columns and no lifecycle. An intention is all lifecycle: open -> asked -> granted / dropped, a partial index for "find the open wants", and a validated status value set. Riding the shared memory_type PG enum would pollute a user-fact taxonomy with an agent-internal control construct and force an irreversible enum add on six consumers.

Enhance-first is still honoured: intention consumes the reusable salience/decay substrate the agent/memory enhancement built (the same salience column shape and the shared apply_salience_decay helper), rather than reinventing it.

Components

intentions table

Partition on agent_id; composite PK (agent_id, intention_id); CAS on date_updated (mirrors memory). Columns: status (PG enum intention_status), content, embedding (pgvector, 1024 wide unless configured -- see below; for dedup on log), salience (NUMERIC(5,4), reuses the decay substrate), last_decayed_at (decay anchor), last_surfaced_at (cooldown anchor), source_memory_id / source_conversation_id (soft-ref provenance, no FK).

Embedding width

intentions.embedding is a plain pgvector vector(N), and N is a deployment setting: agent-memory's width, so one embedding provider serves both corpora. It defaults to 1024, so a host that configures nothing is unaffected. Give the same embedding_dim to threetears.agent.intention.migrations.register(runner, embedding_dim=...), IntentionsCollection(..., embedding_dim=...) and intentions_table(metadata, embedding_dim=...).

• Ceilings. PostgreSQL's pgvector HNSW index takes at most 2,000 dimensions; YugabyteDB's ybhnsw up to 16,000 (no halfvec there, so the column stays vector). register refuses a width below 1 or above 16,000; v001 refuses a width above 2,000 on PostgreSQL before any statement runs.
• Writes are checked. A save whose embedding is not the collection's width is refused with a ValueError naming both widths, before any SQL runs.
• An existing schema keeps its width. await verify_embedding_width(store, embedding_dim) (threetears.agent.intention.embedding_width, also exported from the package root) raises EmbeddingWidthMismatch when intentions.embedding is another width.
• Changing it is an operator action. await resize_embedding_width(store, embedding_dim) drops ix_intentions_embedding_hnsw, alters the column with USING NULL, recreates the index under its own name and options, and returns how many embeddings it cleared. Every want keeps its content and loses its embedding, so re-embed them. Nothing runs it automatically. Run it with writes to these tables stopped: on YugabyteDB the column change is a table rewrite, and writes made while it runs may be lost ("Concurrent DMLs may not be reflected in the new table", yugabyte-db issue 19860).

User isolation is a user_id WHERE clause, not RBAC

Every metallm user shares one agent_id, so the partition isolates nothing and the agent-owner RBAC short-circuit sees every want. user_id is the only boundary: the collection's user-facing reads take user_id as a required parameter, and every consumer read filters on it.

Versioning policy

The package version moves in lockstep with the rest of the 3tears monorepo -- every package tracks the framework git tag.