NorthwoodsSentinel/loam — Repo Appraisal

Overview

What it is — Loam is a self-hosted AI conversation history aggregator that runs entirely on your own Cloudflare account (Worker + D1 + R2), ingesting exports from ChatGPT, Claude, Gemini, Perplexity, and local markdown directories and making them full-text searchable from a single web UI.
Problem — AI context fragments across multiple tools with each conversation starting from zero; Loam provides a persistent, cross-platform memory layer so history from any tool is searchable without sending data to a third-party server.
Who it's for — Power users and builders who operate across multiple AI tools and want a sovereign, queryable record of their complete conversation history on their own infrastructure.
Notable — The "Bring Your Own Cloudflare" sovereignty architecture (no Loam servers ever see your data) and a roadmapped MCP server endpoint that would make the corpus directly queryable by any MCP-aware agent, including a planned counter-thesis surfacing feature.

Verdict

	Rating	Summary
Quality	decent (14/24)	Substantive README and a clean TypeScript/Cloudflare stack, but zero releases, zero forks, no test infrastructure, and a truncated README drag the score to the middle tier.
PAI Relevance	study (0.88)	Sovereign cross-platform AI history substrate with no PAI equivalent — TypeScript/Bun/Cloudflare stack aligns perfectly and the harvest value is high enough to warrant study despite the project's early state.

Quality Assessment

14/24 — stale-risk / adequately-documented / solid

Health: 4/8 (stale-risk)

Failed:

H1: FAIL — no tagged release exists; "Latest Release: none (none)"
H2: FAIL — no release of any kind to assess recency
H6: FAIL — 0 open issues; probe requires >0 AND <100 as a sign of active triage
H8: FAIL — no CI badge or .github/workflows reference anywhere in README

Passed:

H3: PASS — last commit 2026-05-09, 18 days before appraisal date, well within 6 months
H4: PASS — last commit 18 days ago, within the 30-day window
H5: PASS — not archived
H7: PASS — MIT license declared in repo metadata and package.json

Documentation: 5/8 (adequately-documented)

Failed:

D5: FAIL — no heading matching API, Configuration, Options, Reference, Commands, or Parameters; "Architecture" is present but does not match the probe keywords
D7: FAIL — no link to an external docs site, wiki, or /docs directory
D8: FAIL — no Limitations, Caveats, Known Issues, or Trade-offs section; "Project status" paragraph truncates mid-sentence ("but the public release is days old. E")

Passed:

D1: PASS — README is present and very detailed
D2: PASS — README far exceeds 1000 bytes; runs several thousand words
D3: PASS — "Deploy" section contains complete step-by-step bash instructions with wrangler commands
D4: PASS — multiple fenced code blocks appear after ingest and usage headings
D6: PASS — opening line states "The memory substrate for personal AI" within the first 500 characters

Engineering Signals: 5/8 (solid)

Failed:

E4: FAIL — no test script in package.json scripts block; no test infrastructure mentioned anywhere in README
E5: FAIL — 3 stars; well below the 50-star threshold for non-trivial adoption
E7: FAIL — 0 forks

Passed:

E1: PASS — TypeScript is the primary language
E2: PASS — package.json present with version, license, and scripts
E3: PASS — only 3 devDependencies (workers-types, wrangler, typescript); far below the 15-dep CLI ceiling
E6: PASS — 3 stars in approximately 22 days ≈ 4.1 stars/month, above the 2/month threshold
E8: PASS — description is substantive, descriptive, and well above 20 characters

PAI Relevance

Dimension	Score	Assessment
Harvest Value	2	Multi-source AI conversation aggregation with FTS5+BM25 ranking and a planned counter-thesis surfacing feature ("things you've said that contradict what you're saying now") are novel patterns absent from PAI's memory subsystem. The BYOC sovereignty architecture — D1+R2+Worker running entirely on the user's Cloudflare account — is directly applicable to how PAI could structure persistent external memory without centralized hosting.
Integration Readiness	2	TypeScript throughout with explicit `bun ingest/claude.ts` invocation, Cloudflare Workers/D1/R2 deployment — an exact match to PAI's declared TypeScript/Bun/Cloudflare stack. Ingest scripts can be called as PAI skill subprocesses with no adapter layer.
Overlap Risk	1	PAI's `ContextSearch` skill handles session recovery search, and file-based memory covers WORK/LEARNING/KNOWLEDGE directories — these partially overlap with loam's search layer. However, no PAI capability ingests or indexes multi-platform AI conversation exports from ChatGPT, Claude, Gemini, or Perplexity.
Gap Fill	2	PAI has no mechanism for ingesting and querying AI conversation history from external platform exports. Loam fills this gap entirely, providing the cross-session, cross-platform continuity substrate PAI's memory layer currently lacks.

Composite: 0.88

What Next

Capture-to-Knowledge Pipeline at its ingestion schema design stage: Read Loam's src/parsers/ to study how it normalizes four structurally different AI export formats (ChatGPT JSON, Claude markdown, Gemini takeout, Perplexity) into a single turn-level record with consistent source, timestamp, and content fields. Use this normalization contract as the reference schema for your own pipeline's ingestion stage, where different capture types (web clips, voice notes, manual entries) face the identical heterogeneous-inputs-to-one-corpus constraint — you arrive at a field-level schema design with a battle-tested reference rather than reasoning from blank paper.
Any future Cloudflare-native personal tool (including pipeline infrastructure on the Conservancy stack): Study Loam's wrangler.toml and Worker entrypoint to understand how it binds Worker + D1 + R2 into a single wrangler deploy with zero external service dependencies or third-party custody. Treat this as the deployment template for personal data tools on Cloudflare — the BYOC pattern gives you a concrete single-command, sovereignty-preserving reference architecture for anything you build on the stack you already use.
Capture-to-Knowledge Pipeline at its agent-access design stage: Study Loam's planned MCP endpoint and counter-thesis surfacing feature — specifically how it intends to expose an FTS corpus as a ranked, context-aware MCP tool rather than a raw keyword lookup. Use these design decisions as the reference interface for your pipeline's own agent-query layer before implementing it, so you're evaluating a concrete design with documented tradeoffs rather than specifying one from scratch.

Landscape Position

Category: Personal AI & Knowledge

In this category: tinyhumansai--openhuman (excellent/21, skip)

Standing: Loam and openhuman occupy different ends of the personal-ai space — openhuman is an ambitious general-purpose private AI, while loam is a focused sovereign memory substrate with a specific deployment path; they are complementary rather than competitive, though openhuman outscores on quality by a wide margin.

Evidence Base

Density: 7/10 — README present and detailed (though truncated mid-sentence at "Project status"), package.json with full dependency manifest, complete repo metadata (stars, forks, dates, license). Missing: source code files beyond the README, test suite, CI configuration, release tags, and the schema.sql referenced in deploy instructions.

Notes

NorthwoodsSentinel is the same author as NorthwoodsSentinel--brook (fleet monitoring) and NorthwoodsSentinel--meridian-protocol (trust protocol), indicating a coherent multi-project PAI infrastructure effort. The README explicitly references Daniel Miessler's PAi architecture shape as a target use case, making intentional alignment with the PAI ecosystem evident. The roadmapped MCP server endpoint per deployment would make each loam instance directly plug-and-play with MCP-aware agents. The project is 22 days old at time of appraisal — architecture is solid and the README is production-quality, but the implementation surface visible from outside the repo is thin.