- tsconfig.build.json references the nine in-scope packages; each package
is composite and references the workspace packages it depends on.
build:packages is now `tsc -b tsconfig.build.json` instead of a chain
of cd commands.
- Every in-scope package exports its TypeScript entry under a `source`
condition next to the dist/ entries. Vitest, tsx (dev:server, the
Tauri dev launch, the e2e server), the sidecar and launcher bundles and
the web app resolve it, so none of them needs build:packages first or
can pick up a stale dist/. The vitest src aliases remain only for the
excluded packages that have no source export.
- Published packages with a files list (agent, cli, marketplace) pack
src/ too, so the source condition never points at a missing file.
- tsconfig.tests.json and `npm run typecheck:tests` type-check the test
suites of the in-scope packages. They report 426 existing errors, so
the CI step is advisory for now.
- The root test run leaves out the excluded and postponed packages
(hive-mind-*, memory-mcp, wiki-compiler, team-server, waggle-dance,
worker, weaver); they are not built either.
- CI runs the test job on Ubuntu, macOS and Windows with Node 22.23.2
(ADR-0008), and no longer builds the hook runtime and the memory MCP
server.
mem0's delete kept the deleted text in memory_history, and fact extraction
keeps the last messages of each scope in the same file. forget now purges
the memory's history rows, and forgetting a whole scope also drops the
conversation messages mem0 kept for it.
Vectors of different models are not comparable even at the same size; the
store records its embedder in waggle-memory.json and names the mismatch.
Also drop the weaver references from the cli and server tsconfigs.
Add a reference set (100 invented memories across personal memory and two
workspaces, 67 queries: paraphrase, keyword, multi-fact, scope, no-answer)
and a tsx benchmark that loads it into the old engine (MindDB, FrameStore,
HybridSearch, with and without the cross-encoder reranker), the mem0 store
and a plain-cosine baseline, all on Waggle's in-process embedder. It reports
recall@1, recall@5, MRR, latency and misses per engine and query type. The
mem0 half runs on its own (--engines mem0,cosine) once the old engine is
removed. Ollama and the user's own key are optional embedders; the bench
refuses network access to anything else and never reads OPENAI_API_KEY.
Results in docs/oss/research/memory-search-quality.md: mem0 matches the old
engine's chat-recall path (recall@1 0.72, recall@5 0.89 vs 0.90, MRR 0.82
vs 0.83) at a fraction of its latency and is acceptable for v1.
No new dependencies.
- Orchestrator, memory tools, context loader and pattern write-back use a
MemoryStore and a workspace scope instead of MindDB layers; the knowledge
graph tools, cognify, memory linker and feedback handler are gone (graph
hidden in v1, ADR-0002).
- The server opens one mem0 store in <dataDir>/memory/, embedded by the
in-process model on first use, and extracts facts about the user in the
background with the model the user chose (memoryExtraction setting).
- Legacy .mind memories are copied into the store once, in the background;
the legacy files are only read.
- Weaver consolidation, harvest, wiki, memory import, memory lanes, vector
backfill and index reconciliation are removed; agent runs, loops, sub-agent
results, connector fetch and the failed-turn capture write through the
contract. Workspace sessions no longer hold a mind handle.
- FileIndexer and the install audit move to waggle.db; recall date helpers and
the reranker are copied into core.
- The CLI REPL opens the same memory store.
@waggle/memory-contract defines MemoryStore: remember (text, or a
conversation to extract facts from), search, list, count, get, update,
forget, forgetMatching and history over personal and workspace scopes,
with kinds, importance, source, origin and session in place of the old
content prefixes and named sessions. It ships an in-memory store and a
compatibility suite that every implementation runs.
@waggle/memory-mem0 implements the contract on mem0ai 3.3.1 in-process:
SQLite files in one directory, scope as mem0's agent_id, Waggle's own
embedder through mem0's langchain provider, fact extraction through any
OpenAI-compatible endpoint. mem0's telemetry and remote notice download
are switched off before the module loads; a test asserts that no request
is made. Memory survives a restart (tested).
New dependencies: mem0ai 3.3.1 (Apache-2.0, pinned exactly) with its
required peers compromise (MIT) and @cloudflare/workers-types (MIT OR
Apache-2.0); transitive openai (Apache-2.0), axios (MIT) and others under
MIT or BSD-2-Clause.