agentic_memory 0.2.0

Memory for AI agents: conversation, working, long-term and semantic stores, with keyword, embedding and hybrid recall, and summarising history strategies.

Changelog

0.2.0

• recallTool returns untrusted content. Memories are extracted from earlier conversations, which may have contained injected text — recall is how an injection outlives the conversation it arrived in. Behaviour change, so this ships in 0.2.0 only: a tool that is not read-only, called after one of these tools returned, now needs approval.

Breaking

• memoryTools, rememberTool, recallTool and forgetTool take store by name. Every other tool factory in the framework takes its arguments by name; these four were the exception, and callers guessed wrong.

  Migrate automatically:

  dart fix --apply
  

  memoryTools(store)          // before
  memoryTools(store: store)   // after
  

0.1.1

• Shortened the package description to the 60-180 character window pana scores against. Search engines truncate anything longer, so the ten points it withheld were pointing at a real defect: the useful half of the sentence was never being shown.

0.1.0

Initial release of the memory layer.

Added

• Values — MemoryEntry, MemoryKind, MemoryQuery and MemoryHit. One entry shape covers working, conversation, long-term, semantic and shared memory, which differ in scope and retention rather than in mechanics.
• Stores — MemoryStore port; InMemoryMemoryStore with rarity-weighted keyword retrieval blended with importance and recency, deduplication, expiry and value-based eviction; EmbeddedMemoryStore for semantic retrieval with a similarity floor and batched backfill; HybridMemoryStore fusing both rankings with reciprocal rank fusion.
• History strategies — RecallingHistory injects memories relevant to the turn about to be sent; SummarisingHistory condenses older turns and caches, so a steady conversation summarises once per window rather than once per turn.
• Tools — remember, recall and an approval-gated forget, so an agent can manage its own memory.
• Automatic extraction — RememberingAgent runs a separate extraction pass after each successful run, bounded by count and importance, and never lets a failed extraction break a good answer.
• Events — MemoriesRecalled (carrying the statements, so an answer is traceable), MemoriesExtracted, MemoryExtractionFailed and MemoriesPruned.