Plan A: Desktop Embedding Dependency Tiers (Signet-wide) #
Objective
Ship a reliable desktop-first embedding architecture with clear quality tiers while keeping the librarian tied to the user’s existing LLM provider/runtime.
Scope
This plan is for Signet desktop installs and is independent of OpenClaw upstream PR waves.
Tier Model
Librarian rule (all tiers):
- The librarian always uses the user’s existing runtime/provider LLM (OpenClaw/Claude Code/OpenCode/Ollama-hosted chat model, etc.).
- Tiering applies to embeddings only.
Tier definitions:
- Tier 0 (default):
transformers.js
local embeddings- No external process
- No API key required
- Best out-of-box reliability for desktop installs
- Tier 1: Ollama embeddings (example:
`nomic-embed-text`
)- Local model server required
- Better quality than Tier 0 on many setups
- No external API cost
- Tier 3: Remote API embeddings (example: OpenAI
`text-embedding-3-large`
)- API key required
- Highest quality target
- Network dependency and usage cost
Selection + Fallback Behavior
- Respect explicit user config if set.
- If unset, default to Tier 0.
- If Tier 1 is selected but Ollama is unavailable, fail over to Tier 0.
- If Tier 3 is selected but API key/network check fails, fail over to Tier 1 when available, else Tier 0.
Rollout Order
- Ship Tier 0 as stable default for desktop.
- Add Tier 1 with health check + connectivity diagnostics.
- Add Tier 3 with key validation, budget/telemetry guardrails, and clear degraded-mode messaging.
Plan B: OpenClaw Contribution Plan #
Primary Objective
Earn trust with steipete through consistently useful upstream contributions so Signet is a credible option when OpenClaw converges on its long-term memory path.
Context
We reviewed 60+ memory-related OpenClaw PRs and studied the
memory-lancedb
plugin (671 lines) plus `memory-core`
(38 lines).
Current pattern:
- Small, focused, low-risk memory improvements are merged more often
- Large redesigns, dependency-heavy changes, and broad rewrites are usually rejected or stalled
- PRs under ~500 lines with clear test coverage and no migration impact are the highest-probability path
Strategy
Do not lead with “use our plugin.” Lead with concrete improvements that help OpenClaw regardless of whether Signet is adopted.
Guiding posture:
- Helpful and technically precise
- No sales framing inside PRs
- Credit OpenClaw maintainers and existing design decisions
- Keep each PR independently valuable and easy to review
What Stays Differentiated in Signet
- Secret store and vault workflow
- Cross-platform identity portability (
SOUL.md
,IDENTITY.md
,USER.md
) - Full pipeline orchestration (extraction -> decision -> graph -> retention -> summarization)
- Daemon/connectors/harness config synchronization layer
Micro-PR Campaign (ranked by acceptance probability)
Selection criteria: small scope, low coupling, no new dependencies, clear tests, explicit rollback.
Wave 1: Highest probability
PR 1: Content normalization + stronger dedup hash
-
Files:
extensions/memory-lancedb/index.ts -
What: normalize whitespace and casing/punctuation edge cases before hash generation; skip near-duplicates
-
Why: reduces duplicate growth from formatting-only variance
- Size target: ~50 lines
- Migration: none
- Signet source:
platform/daemon/src/content-normalization.ts
PR 2: Importance-aware score calibration
- Files:
extensions/memory-lancedb/index.ts
(MemoryDB.search
) - What: incorporate
importance
into final ranking (for examplescore = (1 / (1 + distance)) * (0.5 + 0.5 * importance)
) - Why: importance is stored but currently does not affect ordering
- Size target: ~20-30 lines
- Migration: none
- Signet source:
platform/core/src/search.ts
(ranking normalization patterns)
**PR 3: Provider availability tracker (ring buffer)**
- Files: new small utility + integration point in embedding call path
- What: fixed-size ring buffer for embedding outcomes
(
success|failure|timeout
) - Why: adds operational visibility without persistent storage changes
- Size target: ~80 lines
- Migration: none
- Signet source:
platform/daemon/src/diagnostics.ts
Wave 2: Medium complexity, still low-risk
**PR 4: Configurable recall threshold ( minScore)**
- Files:
extensions/memory-lancedb/index.ts
- What: expose optional
minScore
inmemory_recall
, validate range, and keep current default when omitted - Why: lets operators tune precision/recall without changing code
- Size target: ~20-40 modified lines
- Migration: none
**PR 5: Token-aware recall budget (optional parameter)**
- Files:
extensions/memory-lancedb/index.ts
- What: optional
contextTokenBudget
; stop adding recalls when budget is reached; default behavior remains current limit-based recall - Why: better context-size control with backward compatibility
- Size target: ~40 lines
- Migration: none
PR 6: Degraded-mode error signaling when embeddings fail
-
Files:
extensions/memory-lancedb/index.ts -
What: on embedding failure, return explicit degraded-mode status and actionable guidance in tool details/logs (instead of generic/no-result behavior)
-
Why: memory availability failures should be diagnosable and not look like “no memories exist”
- Size target: ~40-60 lines
- Migration: none in initial PR scope
- Note: if lexical fallback is desired later, ship it as a separate PR with isolated indexing scope
Wave 3: Submit only after trust is established
PR 7: TTL support for temporary memories
-
Files:
extensions/memory-lancedb/index.ts -
What: optional
expiresAt
, filter expired rows, cleanup path - Size target: ~60 lines
- Migration: likely yes (new column)
- Signet source:
platform/daemon/src/pipeline/retention-worker.ts
**PR 8: Rate-limited repair actions**
- Files: new utility + minimal call-site integration
- What: cooldown + hourly budget for maintenance actions
- Size target: ~100 lines
- Migration: none
- Signet source:
platform/daemon/src/repair-actions.ts
PR 9: Optional reranking with timeout guard
- Files: new utility + search integration
- What: rerank top-N with strict timeout and fallback to original order
- Size target: ~120 lines
- Migration: none
- Signet source:
platform/daemon/src/pipeline/reranker.ts
PR Quality Bar (required for every submission)
Each PR must use OpenClaw’s canonical PR template sections
(.github/pull_request_template.md
) with complete, concrete answers:
- Problem statement
- Scope (what is intentionally out of scope)
- Test plan
- Rollback plan
- Risk notes
- Security impact
- Repro + verification
- Human verification
- Compatibility / migration
- Failure recovery
Minimum test expectations:
- Unit test for new behavior
- Regression test covering the bug/path fixed
- No behavior change in default path unless explicitly intended
- Local validation command succeeds:
pnpm build && pnpm check && pnpm test
Rollback requirement:
- Each PR should be reversible with a small, isolated revert
- Avoid touching multiple subsystems in one PR
AI-assisted disclosure requirement:
- Mark PR as AI-assisted in title or description
- State testing level (untested / lightly tested / fully tested)
- Include prompts/session notes when possible
Signet Plugin Improvements (parallel track)
While contributing upstream, improve @signetai/adapter-openclaw
in
parallel:
- Finalize
openclaw.plugin.json
(kind: "memory"
, completeconfigSchema
, clearuiHints
) - Enforce token budget cap for injected memories
- Verify identity injection (
SOUL.md
/IDENTITY.md
/USER.md
) at session start - Integrate health reporting with OpenClaw doctor flow
- Tight README: 3-command setup + architecture diagram
- Ensure librarian/provider bridge follows OpenClaw provider settings
Community Engagement Plan
GitHub (after Wave 1 PRs are live)
- Comment on issue
#12880
with implementation notes from identity file handling - Comment on PR
#24154
with concrete interoperability learnings - Open one technical discussion: cross-platform identity portability (non-promotional)
X/Twitter (after positive review signal)
- Post concise engineering updates tied to shipped upstream improvements
- Highlight contribution intent first, Signet second
- Keep tone factual and collaborative
Demo (after initial trust is established)
- Same agent persona/context across Claude Code and OpenClaw
- Same memory continuity with no manual reconfiguration
- Short recording showing portability and operational stability
Anti-Patterns
- PRs over ~500 lines
- New core dependencies without prior maintainer buy-in
- Architecture rewrites framed as “replacement”
- Marketing language in technical review threads
- Giving away core Signet differentiators prematurely
Execution Timeline
- Now: finalize Wave 1 specs, exact target functions/files, and tests
- Week 1: submit Wave 1 PRs (small, independent, reversible)
- Week 1-2: continue adapter hardening in parallel
- Week 2-3: submit Wave 2 PRs based on review feedback
- Week 3-4: technical engagement in issues/discussions
- Week 4+: demo and public narrative once upstream trust is visible
Success Metrics
Primary success metric:
- steipete sees Signet as a high-signal contributor and includes us in memory-path shortlist discussions
Leading indicators (measurable):
- Wave 1 merge rate
- Median time to first maintainer response
- Number of requested rework cycles per PR
- Number of follow-up PR invitations or direct maintainer pings
- OpenClaw discussion references to Signet work (technical, not promotional)