Memory and Elixirs

Semantic memory

pgvector similarity search retrieves task-relevant records instead of replaying all history. Context assembly and compaction keep long-running sessions usable.

Context assembly and compaction

Context segments load by priority under a token budget. Soft trimming precedes the pruning cascade, and automatic compaction starts when total context exceeds the model window minus reserved response tokens.

• Provenance records which memories entered a context bundle
• Importance scores and access history guide retention
• Memory flush preserves durable facts before transcript compaction
• Tags support scoped retrieval and selective archival

Portable memory envelopes

Cross-instance exports use signed CBOR envelopes. The import flow verifies an optional Ed25519 signature, checks the BLAKE3 hash, decrypts AES-256-GCM content, and supports selective disclosure and batch transfer.

Elixir lifecycle

Elixirs are durable knowledge objects for skill backups, domain knowledge, context caches, and training data. New knowledge can enter as a draft, receive a quality score from 0–100, then be approved or rejected.

Brewing and injection

Successful work can generate an Elixir draft for review. Approved Elixirs can be injected through semantic search, linked skills, mantra references, or an explicit API request; usage and versions remain traceable.

Elixir interfaces