Fidelis Memory 0.3.0rc1 — records that keep their correction history

Fidelis now lets an agent retain a statement, correct it without erasing the old
version, and retrieve records by validity date through a focused six-tool MCP
surface. This is a pre-release for local macOS/Linux developers with Python 3.10+
and Ollama. See the user-fit matrix.

python3 -m pip install "fidelis-memory==0.3.0rc1"

For the Gemini extension, use the v0.3.0rc1 ref. The Gemini and Claude plugin
manifests pin the matching package.

Changes

• Time-aware records with stable IDs, validity intervals, explicit supersession,
  historical recall, and correction browsing.
• Truthful write dispositions for stored, duplicate, rejected, and queued writes.
• Six MCP tools: recall, store, correct, get, recent, and health.
• Full memory text in MCP recall, recent lists, and get correction chains; the
  old silent preview cutoffs are removed.
• CLI store and add use verbatim storage; add --extract explicitly selects
  the legacy extraction path.
• Fast vector retrieval by default; explicit thorough zero-LLM hybrid recall.
• Public lazy startup, service collision protection, port overrides, and the
  existing five-client installation routes are retained.

Upgrade and rollback

Back up your configured store and queue and keep the prior package environment.
Stop the old service, install the pinned candidate, then restart with the same
store, collection, namespace, and embedding model. Do not change the embedding
model within an existing collection. Inspect health and a known record before
using the candidate for new writes.

MCP tool names change: use fidelis_recall in place of fidelis_query and update
instructions that call fidelis_orient or cogito_* tools. Restart clients so
cached tool lists refresh. The HTTP legacy retrieval routes remain available.

For rollback, stop the candidate and restore the previous environment and the
pre-upgrade store/queue backup. Do not assume the older package understands new
correction metadata. Preserve the candidate store separately if it has new writes.

Evidence and limits

Release evidence covers an isolated fresh-wheel install, HTTP and real MCP
store/recall/correction/as-of flows, restart persistence, unavailable-Ollama
startup, port parity, and the six-tool list, plus the repository test/lint gates.
Fresh default-path retrieval evaluation is recorded in the
methodology and
results, separately from historical
hybrid and QA results. No historical headline score is claimed for this redesign.

Full LLM/QA evaluation and non-blocking documentation/optimization remain
post-release work. Semantic retrieval is fallible; dates and links describe
recorded assertions, not independently verified truth. Secret screening is
pattern-based. See the README limitations.

The default evaluation completed against its recorded unchanged source snapshot. Subsequent release fixes changed only the package docstring and the older-protocol queued-write acknowledgement in _tool_store; evaluated retrieval code is unchanged. The optional thorough comparison completed in 4m43s, but its source-change flag is retained and it is not used as a headline release benchmark.

Hybrid retrieval tuning is deferred to follow-up work; fast zero-LLM recall remains the default for this architectural pre-release.