v0.3.0

MemoryWhale v0.3.0 — agent memory layer, multi-machine sharing, secre…

MemoryWhale v0.2.1 — semantic recall

Recall now understands meaning, not just keywords.

New since v0.2.0

• Semantic recall (#24) — retrieval can use local Ollama embeddings (nomic-embed-text, 768-d) for cosine similarity, with a lexical fallback when Ollama isn't running, so the app stays zero-setup. Example: "where do we keep our data?" now surfaces a "use mw-core for storage" memory that shares no literal words. The Recall input is debounced so semantic mode doesn't re-embed on every keystroke.
• Landing site on GitHub Pages (#23) — https://wuisabel-gif.github.io/MemWhale/ — with the v0.2.0 banner and cross-links to the sibling project Delphin.

Enabling semantic mode

ollama pull nomic-embed-text     # then keep the Ollama server running

Without it, Recall transparently falls back to lexical matching — no breakage.

Unchanged from v0.2.0

Explainable retrieval (ranked memories with "retrieved because…" per-signal reasons) over your notes, terminal commands, and agent conversations.

MemoryWhale v0.2.0 — the app renders now + explainable recall

Headline

The desktop app actually renders its UI now — and it can explain its memory.

Fixed (the big one)

• The Vite entry (src-tauri/index.html) was a static landing page with no #root/main.tsx, so the React UI (knowledge galaxy, panels) never mounted — the app only ever showed a landing page. Restored a proper Vite React shell (the old page is preserved as landing.html). The build now bundles the full app (3 → 2,133 modules).

New: explainable recall

• Recall panel (left sidebar): ask your memory a question and get ranked hits, each with a "retrieved because…" breakdown — click a hit to see the per-signal bars.
• Retrieval blends interpretable signals: similarity · recency · importance · reinforcement · task-relevance, each with a human-readable reason.
• Searches across everything MemoryWhale remembers: notes, terminal command runs, and agent conversation turns (new agent_turns table).
• New mw-memory crate — a pluggable MemoryEngine: local-first semantic embeddings via Ollama (lexical fallback offline), plus a MemPalace adapter seam.

New: direction

• VISION.md — MemoryWhale as an inspectable memory OS (capture → filter → consolidate → link → retrieve → explain → forget).

Notes

• Similarity is lexical by default; install Ollama (ollama pull nomic-embed-text) for semantic recall.
• agent_turns is populated by external writers (e.g. Delphin / hooks); it may be empty on a fresh install.

Verified

cargo check ✓ · frontend tsc && vite build ✓ · mw-memory unit tests 10/10 ✓ · recall smoke-tested against a real database (correct top hit with reasons).

MemoryWhale v0.1.0

MemoryWhale's first release — a local-first, Rust/Tauri terminal-memory system that turns your commands, errors, fixes, and whole shell sessions into a durable, searchable, visual memory. Everything stays on your machine; nothing is uploaded.

Highlights

Record

• mw — record a whole interactive shell session (every command + real output); mw list / mw show <id>, and mw global on|off to auto-record every new terminal.
• mw-remember — log a single command run (args, exit code, stdout/stderr, notes).
• mw-recover — rescue an interrupted session whose recording was cut off before it saved (also auto-runs when the dashboard starts).

View

• mw-view <id> — open one memory as a friendly local web page.
• mw-serve — a local web dashboard (works headless / over the LAN, ideal for a Jetson): browse command runs and sessions, open detail pages with suggested next steps mined from your history, an interactive knowledge graph (commands sized by frequency, shared arguments as bridges), and project grouping that stitches multi-terminal work into one timeline via a project:<name> tag.

For agents

• A Claude Code skill (.claude/skills/memorywhale/) and AGENTS.md (Codex) that teach AI coding agents to recall prior attempts before debugging and record what they do.

Install

Build from source (see SOP.md and DEBUG.md):

cd src-tauri
cargo build --release --bin mw --bin mw-remember --bin mw-serve --bin mw-view --bin mw-recover
mkdir -p ~/.local/bin && cp target/release/{mw,mw-remember,mw-serve,mw-view,mw-recover} ~/.local/bin/
# macOS only: codesign --force --sign - ~/.local/bin/mw*

Data

Local SQLite at <data_local>/MemoryWhale/memorywhale.sqlite3 (~/.local/share on Linux/Jetson, ~/Library/Application Support on macOS).

See README.md for the full overview and SOP.md for day-to-day usage.