Release 7.0.5 #43

Merged

ric merged 10 commits from v7.0.x into main 2026-10-29 11:01:10 +00:00

Conversation 0 Commits 10 Files changed 45 +1025 -93

ric commented 2026-10-29 10:28:16 +00:00

Owner

• Copy link
• Copy Markdown

OmniMem 7.0.5, tagged v7.0.5 on v7.0.x. The full notes are in CHANGELOG.md.

• Current Claude models work for fact extraction and the other Claude calls: OmniMem now reads the first text block of a reply (GitHub #4). Claude Haiku 5.5 is the default model.
• Memories imported into a running server's database are found without a restart (#5).
• A refused Host gets 421 on /mcp, before the token check: MCP_ALLOWED_HOSTS=* is now accepted, there's a startup hint, and the docs are fixed (#6).
• Recall results add similarity (raw similarity, beside the score they're ranked by), and a forget preview adds a hint saying how to confirm (#7).
• RECALL_EXPERIENCE_SPAN (opt-in): experience is added to the score instead of multiplying it, so it can't outrank a clearly better match (#8).
• EMBEDDING_QUERY_PREFIX and EMBEDDING_PASSAGE_PREFIX: for models such as bge and e5, which expect a prefix (#9).
• Extracted facts follow their source through the lifecycle and carry as_of, and preferences are stricter (#10).
• CONTRADICTION_MODEL and RSS_SUMMARY_MODEL: both are configurable from the environment and the settings window (#11).

🤖 Generated with Claude Code

https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

OmniMem 7.0.5, tagged `v7.0.5` on `v7.0.x`. The full notes are in CHANGELOG.md. - **Current Claude models work for fact extraction and the other Claude calls**: OmniMem now reads the first text block of a reply (GitHub #4). Claude Haiku 5.5 is the default model. - **Memories imported into a running server's database are found without a restart** (#5). - **A refused Host gets 421 on `/mcp`, before the token check**: `MCP_ALLOWED_HOSTS=*` is now accepted, there's a startup hint, and the docs are fixed (#6). - **Recall results add `similarity`** (raw similarity, beside the score they're ranked by), and a `forget` preview adds a `hint` saying how to confirm (#7). - **`RECALL_EXPERIENCE_SPAN`** (opt-in): experience is added to the score instead of multiplying it, so it can't outrank a clearly better match (#8). - **`EMBEDDING_QUERY_PREFIX` and `EMBEDDING_PASSAGE_PREFIX`**: for models such as bge and e5, which expect a prefix (#9). - **Extracted facts follow their source through the lifecycle and carry `as_of`**, and preferences are stricter (#10). - **`CONTRADICTION_MODEL` and `RSS_SUMMARY_MODEL`**: both are configurable from the environment and the settings window (#11). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

ric added 10 commits 2026-10-29 10:28:16 +00:00

fix: Claude replies are read from their first text block, and Haiku 5.5 is the default ... 9851f2c963

Newer models can open a reply with a block that isn't text, and reading
content[0] as 6.x did failed every fact extraction with "no text in the
first content block". The default model for extraction, query expansion,
contradiction checks and RSS summaries moves to claude-haiku-5-5.
Version 7.0.5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

feat: RECALL_EXPERIENCE_SPAN makes the experience boost a bounded nudge ... 708303fae2

Recall multiplies by the experience weight however relevant a memory is,
so a boosted loose match outranks a near-exact one: a user's 36-query
benchmark found 20 targets ranked first on raw similarity and 11 through
recall(). Above 0, the weight adds (weight - 1) * span instead, which
settles near-ties without beating clearly better matches. 0, the default,
keeps 6.x's multiplier.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

docs: the 7.0.5 notes link GitHub #4 and #8 [skip ci] ... 47d3b35457

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

fix: a refused Host gets 421 on /mcp before the token check, and the docs say how hosts match ... a6ad0019d9

/mcp answered a Host that isn't allowed with rmcp's 403 while the docs
said 421, and the token check ran first, so a probe without a token saw
401 and looked healthy. OmniMem's guard now sits in front of /mcp ahead
of authentication, logs the setting to change, and accepts `*`. The
server warns at startup when it listens everywhere with no host names
allowed. Docs: how ports and `*` match, 3.x imports, the git pull trap
for 6.x clones of main, the empty priming SSE event. The search help
text and the dead Anthropic feed are fixed too. GitHub #6.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

feat: recall results carry raw similarity, and a forget preview says how to confirm ... af40bc5048

`score` is the ranked score while the relevance floor and weak band read
raw similarity, so results now show both. GitHub #7.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

fix: memories another process writes are found without restarting the server ... d7735045dd

The vector matrix is loaded once at start, so an `omnimem import` beside a
running server stayed invisible to recall until a restart. The store now
checks PRAGMA data_version before searching and reloads the matrix when
another connection has committed. GitHub #5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

feat: query and passage prefixes for embedding models that expect them ... 79aba3e4ed

EMBEDDING_QUERY_PREFIX goes before every search query and
EMBEDDING_PASSAGE_PREFIX before everything stored, so models such as
bge-*-en-v1.5 and e5 can be used as trained. TextEmbedder gains
embed_queries (defaulting to embed_texts), and the engine's searches use
it. Unset by default, so MiniLM's vectors don't move; a bge or e5 model
without prefixes gets a startup warning. Also corrects the --no-embed
hint, which pointed at reindex (that reloads vectors, it doesn't embed).
GitHub #9.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

feat: extracted facts follow their source, carry an as-of date, and preferences mean rules ... 9f541ed611

Deprioritising, archiving or forgetting a memory now does the same to the
facts and preferences extracted from it, and reinstating it restores the
ones it took, as they were. Facts store `as_of` (their source's
created_at) and recall shows it as a date. The extraction prompt files
statuses, tasks and plans as facts, never preferences. A deprioritise by
query tolerates a fact its source has already taken. GitHub #10.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

feat: CONTRADICTION_MODEL and RSS_SUMMARY_MODEL pick the model for those Claude calls ... 006360f4b0

The contradiction check and RSS summaries (with digest item extraction)
were fixed to Haiku, as in 6.x. Both now read a model from the
environment, omnimem.env or the settings window's Configuration page,
defaulting to claude-haiku-5-5. Closes #11 on release.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

release: OmniMem 7.0.5, with configurable models, prefixes and facts that follow their source ... 858fa12671

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V9cZqempzHD2h4Xyjm3gFL

ric merged commit 3790d5b752 into main 2026-10-29 11:01:10 +00:00

ric referenced this pull request from a commit 2026-10-29 11:01:11 +00:00

Release 7.0.5 (#43)

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

ric/omnimem!43

No description provided.