0.3.0

Added

• document memory type (long-term, importance 0.6): retain(type="document")
  stores the body through a DocumentStorePort, indexes its token chunks for
  hybrid search and keeps a summary record; get returns the summary, recall
  resolves chunk hits to the document and never lists chunk records (ADR-009).
• ask(ref_id, question, mode="auto") MCP tool and MemoryService.ask:
  answer a question about one memory without loading it; mode is
  direct | rag | rlm | auto, passed to the reasoner unchanged.
• Ports DocumentStorePort and ReasonerPort in core/ports.py; Answer
  model; built-in LocalDocumentStore (ChromaDB) and LLM-free LocalReasoner
  (dense + BM25 + RRF over chunks, mode="rag").
• RLM Studio adapter memkernel.adapters.rlmstudio.RLMStudioReasoner
  (MEMKERNEL_REASONER=rlmstudio, lazy import, falls back to the local
  reasoner with a warning when rlmstudio is not installed).
• consolidate(synthesize=True): with an external reasoner, drafts long-term
  statements from the candidates under "suggested".
• memkernel.lifecycle.MemoryLifecycle with on_session_start,
  on_pre_compact, on_post_compact, on_session_end; the hook templates
  are thin wrappers over it.
• Settings MEMKERNEL_DOCUMENT_STORE, MEMKERNEL_DOCUMENT_SUMMARY_CHARS,
  MEMKERNEL_REASONER, MEMKERNEL_ASK_TOP_K; split_text / TokenChunker
  in retrieval/chunking.py.
• Port additions for adapter authors: VectorStorePort.count(where=None)
  (filtered, ids-only count) and DocumentStorePort.indexed(doc_id) (number
  of indexed chunks); both must be implemented by custom backends.

Changed

• consolidate reports always carry a suggested list (empty unless
  synthesize=True with an external reasoner).
• templates/claude-memory-instructions.md tells the agent to use
  retain(type="document") + ask() for anything longer than a screen.
• Release workflow uses actions/upload-artifact@v5 and
  actions/download-artifact@v6 (Node 24).

Removed

• The forget_expired MCP tool (purging stays inside consolidate; the
  MemoryService.forget_expired method remains). The server exposes nine
  tools: retain, recall, get, forget, promote, consolidate, ask, wiki_export, wiki_import.

0.2.0

0.2.0 - 2026-10-11
Added
Tiered memory model: working, short_term, long_term tiers derived from the memory type; new types fact, procedure, observation, task (ADR-007).
Per-memory importance, tags, access_count, last_accessed, supersedes and promoted_from metadata. Records written by 0.1.x load unchanged with derived defaults.
Recall re-scoring by relevance + recency + importance with configurable weights; recency decays from last access; superseded memories are hidden.
Access-reinforced retention for short-term and working memories (MEMKERNEL_ACCESS_BOOST, MEMKERNEL_OBSERVATION_TTL_DAYS, MEMKERNEL_TASK_TTL_HOURS, MEMKERNEL_RECENCY_HALF_LIFE_DAYS).
MemoryService.promote, consolidate, forget_expired, list_memories.
MCP tools promote, consolidate, forget_expired, wiki_export, wiki_import; recall gains a tier filter; retain gains importance, tags, supersedes.
LLM-wiki bridge (memkernel.wiki): compile long-term memories into draft markdown pages and index curated pages as memories (ADR-008); MEMKERNEL_WIKI_DIR setting.
memkernel.factory.build_service() as the supported Python entry point; templates/hooks/session-start.py.
Design docs: architecture, literature review, sibling-project map, ADRs 007–009, licensing review, roadmap, harness integration plan.
Release workflow using PyPI trusted publishing.
Changed
License changed from MIT to Apache-2.0.
recall and get record an access (pass touch=False to avoid it).
/stats reports a per-tier breakdown.
Hook templates build the service through build_service() and surface promotion candidates after compaction.
CI frees runner disk before installing dependencies; actions moved to checkout@v5 / setup-python@v6.