NorthwoodsSentinel/loam

*The memory substrate for personal AI. Bring Your Own Cloudflare. The soil layer beneath whatever you build on top.

TypeScript3 starsPersonal AI MemoryGitHub

Standalone Assessment

Maturity: 2/5

Explicitly alpha — the README itself says "the public release is days old." Created 2026-05-05, last commit 2026-05-09: six days of public existence at time of appraisal. No releases tagged, no changelog, no versioned milestones. The schema includes a migration file (0001_provenance_and_sensitivity.sql) which is a good sign that the author is thinking about schema evolution, but there is no evidence of real-world stress-testing beyond "running on real personal corpora at thousands of entries." Zero open issues likely reflects zero external users, not issue triage quality.

Documentation: 4/5

Remarkably thorough for a six-day-old project. The README covers motivation, architecture (ASCII diagram), step-by-step Cloudflare deployment with exact CLI commands, per-platform ingest instructions (Claude, ChatGPT, Gemini, Perplexity), security posture, and a roadmap. The philosophy section ("Loam, n.") sets clear design intent. No docs site, no API reference, no contribution guide — points deducted there — but for standing up and using the tool the README is nearly self-contained.

Code Quality: 3/5

TypeScript throughout, which is appropriate for a Cloudflare Worker project. Dependency footprint is admirably minimal: only @cloudflare/workers-types, wrangler, and typescript as devDependencies — no runtime bloat. Ingest scripts use bun (reasonable for local scripting). FTS5 + BM25 on D1 is a sensible, fast search choice for individual-scale corpora. No tests are visible, no CI pipeline mentioned. Architecture is clean (Worker / D1 / R2 separation, parameterized D1 REST calls from ingest scripts). Cannot assess internal code structure beyond what the README reveals, which limits this score.

Maintenance: 2/5

Four commits in four days is active, but the observation window is too narrow to establish a maintenance pattern. Single apparent author (NorthwoodsSentinel). No PRs, no issue responses to evaluate, no community channel mentioned. The roadmap is concrete (Vectorize, MCP, proactive surfacing, counter-thesis), which suggests intentional development direction, but a solo project this new carries real abandonment risk.

Adoption: 1/5

3 stars, 0 forks, no downstream dependents visible, no package published to npm. Entirely expected for a six-day-old personal tool, but the signal is what it is. Star trajectory is unmeasurable over such a short window.

Overall: 2.2/5

Competitive Positioning

Category: Personal AI Memory Known alternatives in vault: None — this is the first appraised repo; no prior entries exist in any category. Differentiation: Loam's core differentiator is the combination of (1) multi-platform AI export ingestion with concrete pipelines for Claude, ChatGPT, Gemini, and Perplexity already built or committed, (2) strict "Bring Your Own Cloudflare" sovereignty — the authors provably never touch user data, (3) FTS5/BM25 full-text search rather than vector-only retrieval, giving sub-100ms latency at corpus scale without embedding costs, and (4) a planned MCP server endpoint that would make it a plug-in memory layer for any MCP-aware AI client. Conceptual alternatives in the broader ecosystem (Rewind, MemGPT/Letta, personal Chroma/Qdrant deployments) either require proprietary infrastructure, focus on in-session memory rather than archival search, or lack the cross-platform export ingestion story. Gap or crowd: Fills a clear gap. The vault has no other personal AI memory or conversation-archival repos. This is the first entry in what is likely a high-priority category for any PAI build.

PAI Fit

Score: 5/5 Harvestable: The FTS5+BM25 D1 schema and search query patterns are directly reusable. The ingest scripts (claude.ts, chatgpt.ts, gemini.ts, files.ts) encode the normalization logic for every major AI export format — the conversation/message data model they produce is a clean substrate. The files.ts source-tagged markdown ingest pattern is immediately applicable to any personal knowledge vault directory. The bearer-token Cloudflare Worker auth pattern is a reusable micro-pattern. Integration path: Deploy as-is to Cloudflare Workers (documented 5-minute path), run ingest scripts against existing AI exports, query via REST search API from any tool or agent. The upcoming MCP endpoint would make it a first-class PAI tool node — any MCP-aware agent (Claude Code, Cursor, custom agents) could query the memory substrate directly without bespoke integration. Until MCP lands, the REST API is queryable from a PAI skill with minimal glue code. Overlap with existing: No repos currently in the vault overlap. If the vault later acquires a vector-search or retrieval-augmented-generation repo, there will be partial overlap on the "query past context" use case, but Loam's ingestion and sovereignty posture would remain distinct. Adoption cost: Trivial to moderate. Deployment is well-documented and infrastructure cost is effectively zero at personal scale on Cloudflare's free tier. The main cost is running ingest scripts against AI platform exports (one-time effort per platform, then periodic re-ingestion). No code modifications required for the baseline use case. Integration into a PAI skill layer would require writing a thin REST client wrapper — moderate if MCP is not yet available, trivial once the MCP endpoint ships.

Notes

Loam is the most purpose-aligned repo a PAI vault could acquire: it is explicitly designed as the memory substrate layer for personal AI, built on sovereign infrastructure, with a roadmap (MCP, proactive surfacing, counter-thesis detection) that maps almost point-for-point onto PAI architectural needs. The weakness is purely one of age — six days of public existence means no track record, no community, and real abandonment risk from a solo author. The architecture is sound and the documentation is unusually good for the age. Recommended for immediate watch-list and trial deployment against personal AI exports. Re-appraise at 90 days or on first tagged release, whichever comes first, to assess whether the roadmap items (Vectorize, MCP) are landing and whether the author is still active.