Changelog

What's new in Phorvec, newest first.

v0.1.9 — 2026-09-01

A data-loss fix. If you run more than one Phorvec process against the same project, upgrade.

  • Concurrent writers no longer destroy each other's data — a handle loaded a project's cell into memory once and wrote that snapshot back on every save, so anything another process committed in the meantime was silently discarded. Two Claude Code windows, or an MCP server plus a Python script, on the same project could each revert the other's work. Writes are now a locked read-modify-write that merges per key, so both sides' keys survive. Existing .avdb files load unchanged.
  • Long-lived servers stop serving stale reads — a cached handle now refreshes when the file underneath it has moved on, instead of answering from a snapshot that could be days old.
  • conflict_check no longer reports false criticals — two decisions that agreed could be flagged as a Critical contradiction because of unrelated wording elsewhere in the text. Negation and opposition are now scoped to the term actually in dispute. project is no longer required.
  • Em-dashes and emoji no longer crash tool calls — content previews truncated on byte boundaries, so a multi-byte character near the cut panicked the request. Affected conflict_check, team_feed, team_search, context_retrieve, context_optimize, context_store and memory_recall.
  • The embedded Engine reads config.toml — pip and embedded users were silently getting compiled defaults, with the whole config surface unreachable. ⚠️ If you already keep a config.toml for your MCP server, your embedded engines will now start honouring it — check [embedding] and [quantization] before upgrading, as changing those invalidates stored vectors or needs a migration. Pass config_path=False to keep the old behaviour.

v0.1.8 — 2026-08-01

  • Delete a memory by id — the new memory_delete tool removes a memory-store item: soft by default (hidden from memory_recall immediately, recoverable), or hard: true to purge it permanently. From Python: memory.delete(id). It's the counterpart to memory_forget, which only affects context items.
  • context_retrieve filters are honored again — a semantic query-cache bug could return a stale, unfiltered response when the same (or near-worded) query was re-run with different filters / top_k / as_of. The cache now includes those parameters in its key, so filtered queries return filtered results.

v0.1.6 — 2026-07-27

  • Re-ingesting a project no longer stacks duplicate memories. context_store gains a replace flag: pass replace: true to clear a project's existing context before writing, so deterministic re-runs don't silently pile up duplicates. The clear is atomic with the write and branch-aware; for a standalone wipe, use context_clear.

v0.1.5 — 2026-07-19

Going paid no longer means reinstalling — one command swaps the wheel in place.

  • phorvec upgrade <YOUR-KEY> — one command swaps a free-wheel install for the paid wheel from your licensed index and unlocks paid tools (restart your MCP server / Python process after upgrading). Brand-new installs can point pip install at the licensed index directly. Details →
  • Free installs holding a paid license get nudged — if a paid key is already activated on a free wheel, phorvec now tells you to run phorvec upgrade instead of silently running the free tool set.

Python package only. The standalone MCP-server binaries on /download remain at v0.1.4 — pip install phorvec is the current release.

v0.1.4 — 2026-07-18

The full memory engine — teams and all — now runs in-process from pip install phorvec.

  • Full tool parity in the Python library — every tool the MCP server exposes (147 on the paid tier) is now callable straight from the pip package, with the same license gating. No server to host.
  • In-process multi-agent teams — form a team and act as any member from a single Python program: shared team feed, cross-member conflict detection, and cross-agent insight discovery, all via the new phorvec.Team API (Team.as_member(...)).
  • Fixed team-feed membership — a member-id derivation bug could hide teammates' shared items from the feed; teammates now show up reliably.

v0.1.3 — 2026-07-07

Your sessions now survive /compact — and your memory is portable.

  • Session continuity for Claude Code — one command (phorvec init --tool claude-code --hooks) installs hooks that save a distilled summary of your session before compaction or exit, and restore the key decisions into your next session automatically. Fully local, and hardened so a hook failure can never break a session. In our deterministic eval, 5/5 seeded decisions survive compaction (8/10 at N=10) — baseline without it is 0.
  • Import your memory from anywhere — phorvec ingest brings in ChatGPT exports, CLAUDE.md/MEMORY.md files, mem0 exports, and Obsidian vaults.
  • Open .avdb format spec — the portable agent-memory archive is now a documented open format, so your memory is never locked in. phorvec export --quantize int8 shrinks archives ~4x.
  • Team servers, one line per engineer — phorvec init --remote https://memory.internal:8080/mcp --remote-token pv_<key> points Claude Code at your self-hosted team memory.
  • Leaner MCP surface — new installs default to a minimal 18-tool preset (--preset minimal|standard|full).
  • Hybrid retrieval fix — the default hybrid mode no longer underperforms on verbatim-stored content (RRF fusion bug; benchmark restored to parity with vector mode).

v0.1.2 — 2026-07-03

Phorvec is now on PyPI — pip install phorvec.

The memory engine behind the MCP server is now a Python library you can drop straight into your own agents.

  • pip install phorvec — the full engine, in-process, free. Python guide →
  • Paid tiers upgrade in place — install from the licensed index with your key to unlock every tool, including shared team memory. No code changes. Pricing →
  • Free community signup — grab a key for release notes and our Discord. Community stays free and needs no key to run.

v0.1.1 — 2026-06-20

  • One-command updates — phorvec update upgrades you to the latest release in place.

v0.1.0 — 2026-05-23

  • Phorvec launches — persistent vector memory for AI agents: one file per agent, local-first, MCP-native.
  • Hybrid search — semantic vectors plus keyword matching, with context management, memory consolidation, and multi-agent tools.
  • Signed binaries for macOS, Linux, and Windows.