Conversation
…on DSL - Move kndl-memory-pack/kndl-memory-mcp/ → packages/kndl-memory/ (@kndl/memory 2.0.0-alpha.1) - Move kndl-memory-pack/kndl-memory/ → skills/kndl-memory/ (drop-in Claude Skill bundle) - Delete packages/python/, packages/mcp-server/ (Python DSL retired) - Delete spec/ (SPECIFICATION.md and kndl.ebnf — no DSL = no grammar) - Delete kndl-memory-pack/ wrapper directory - CI: drop python + mcp-server jobs, add node-test job - Build verified: dist/ generates clean, both binaries (kndl, kndl-memory-mcp) present
…ackends) - src/types.ts: FactStore async interface (assertFact, supersedeFact, query, contradictions, provenanceChain, list, show, close) - src/core.ts: pure logic only — decay math, time utils, fact construction, applyQuery, findContradictions, buildProvenanceChain (no FS/DB imports) - src/stores/fs.ts: FsFactStore — filesystem backend (Anthropic Memory mount) - src/stores/sqlite.ts: SqliteFactStore — better-sqlite3, WAL, indexed columns - src/stores/duckdb.ts: DuckDbFactStore — @duckdb/node-api, columnar - src/stores/supabase.ts: SupabaseFactStore — @supabase/supabase-js, RLS - src/stores/index.ts: makeStore() factory dispatching on KNDL_STORAGE env var - cli.ts + server.ts updated to use makeStore() - tests/stores.test.ts: 25 tests across FsFactStore + SqliteFactStore on the loan-decision corpus (assertFact, list, query, contradictions, show, supersedeFact, provenanceChain) + decay math unit tests
- src/notify.ts:
- NotifyingStore: wraps any FactStore, emits "change" EventEmitter events
on every assertFact / supersedeFact call (in-process notification)
- attachFsWatcher: chokidar watches facts/ dir for cross-process FS writes
- attachSqlitePoller: polls kndl_changes table for cross-process SQLite writes
- logSqliteChange: helper for SqliteFactStore to write change-log entries
- SubscriptionRegistry: filter-based subscription map with matches() ranking
- server.ts:
- Uses NotifyingStore; broadcasts notifications/resources/updated to all
active sessions when a subscribed fact changes
- New tools: subscribe, unsubscribe, list_subscriptions
- New resource: kndl://fact/{id} — live snapshot with effective_confidence
- --http flag: StreamableHTTPServerTransport on port $PORT (default 8000)
with per-session transport map (fixes 406/400 errors)
- stdio mode unchanged; both modes register activeSessions for broadcast
Each bundle exercises: confidence, decay, supersession, contradiction, derivedFrom, and classification. All 42 facts pass JSON + Fact-shape validation. loan-decision/ 7 facts — credit score supersession, employment, financial distress iot-sensor/ 6 facts — temperature decay/supersession, occupancy, HVAC alert (derivedFrom) personal-memory/ 5 facts — role supersession, PII classification, preference (derivedFrom) threat-intel/ 5 facts — IOC decay, false-positive retraction (negated), contradiction clinical/ 5 facts — PHI, diagnosis supersession, negated allergy, Rx (derivedFrom) legal-ediscovery/ 4 facts — testimony supersession, contradicting witness, document auth scientific-lab/ 5 facts — replicate measurements, outlier retraction, consensus (derivedFrom) ai-evals/ 5 facts — benchmark decay, regression contradiction, capability (derivedFrom)
- src/remote/types.ts: MemoryStoreItem, MemoryStoreClient interface,
RemoteConfig, SyncResult — typed contracts for the remote layer
- src/remote/anthropic.ts:
- AnthropicMemoryClient: REST client for Memory Stores beta API using
fetch() (no SDK dep). Exponential backoff on 429/529. Pagination via
cursor. NOTE: watch_memory_store deferred pending watermark verification
(plan Q7 — if /since param confirmed, re-enable)
- FakeMemoryStoreClient: in-process fake for tests; seed + paginate
- src/remote/sync.ts: pull() driver — translates Memory Store items to
KNDL Facts; idempotent on source URI + content-hash tag; supersedes on
content change; contradiction detection after each batch.
push() → explicit NotImplementedError (v2.1)
- src/remote/config.ts: ~/.kndl/remotes.json read/write; KNDL_REMOTE_STORES
env var shorthand; addRemote / removeRemote / getRemote helpers
- src/types.ts: add tags?: string[] to FactInput
- src/core.ts: buildFact now copies input.tags onto the Fact
- server.ts: sync_memory_store + list_memory_stores tools (gated on
ANTHROPIC_API_KEY); watch_memory_store NOT shipped (pending Q7)
- cli.ts: kndl remote add|pull|ls|rm subcommand
- tests/remote.test.ts: 11 loopback tests using FakeMemoryStoreClient —
no real API calls; covers pull, idempotency, supersession on change,
push stub, pagination, SQLite backend
…aceholder)
eval/questions.json (rewritten):
- 33 binary-scored questions across 8 domains × 7 archetypes
- 8 cross-domain (loan-decision, unchanged from v1)
- 4 IoT sensor, 3 personal-memory, 4 threat-intel, 4 clinical,
3 legal-ediscovery, 3 scientific-lab, 3 ai-evals
- 1 cross-runtime (manual; requires watch_memory_store + managed agent)
- Threshold: ≥70% KNDL beats vanilla required to ship v2.0
eval/runner.ts:
- Drives Claude in vanilla mode (facts in system prompt)
- Claude-as-judge auto-scores each answer binary PASS/FAIL
- Outputs results.json with pass rate, per-archetype breakdown, verdict
- MCP mode documented but deferred (requires running server during eval)
- Run: ANTHROPIC_API_KEY=... tsx eval/runner.ts [--scenario <id>] [--out <path>]
eval/results.json: placeholder; populated by running eval/runner.ts
Route changes (v2.md §9):
Deleted: /spec, /spec/full, /workflow (SpecPage, SpecFullPage, WorkflowPage)
New: /protocol, /skill, /examples, /eval
Rebuilt: /explorer (fact bundle viewer), / (landing), /mcp (TS install)
Redirects: /spec → /protocol, /spec/full → /protocol, /workflow → /skill
New pages:
ProtocolPage — annotated Fact shape, 21-field reference, decay formula
SkillPage — 6 workflow steps, trust thresholds, reasoning rules, decay table
ExamplesPage — 8 domain cards with fact counts and feature badges
EvalPage — scoreboard (NOT_RUN state + results state)
Rebuilt pages:
ExplorerPage — domain selector, fact cards, effective-confidence bar,
supersession greying, SPO rows, classification badges
LandingPage — WHERE/WHAT/HOW three-layer model, problem statement,
fact code block, 6 pillars, 6 use cases, CTA row
McpPage — @kndl/memory TypeScript install, 9 tools, remote sync CLI
Data:
src/data/examples.ts — all 42 facts across 8 domains embedded as TS constants
Infra:
Nav updated (7 links), sitemap.xml (7 routes), llms.txt rewritten
Build: 367KB JS / 112KB gzip, 6 prerendered route shells
… release Migration tool (kndl migrate): - Reads v1 SQLite database (kndl_nodes/kndl_edges/kndl_intents) - Maps each Node → Fact (subject=node:id, predicate=isa, object=type_name) - Maps each Edge → Fact (subject=source, predicate=edge_type, object=target) - Maps each Intent → Fact (@type=Action) - --dry-run flag for preview; idempotent (skips existing files) - kndl migrate --from sqlite:./kndl-v1.db --to ./memory Schema publish: - website/public/context/v1.jsonld — JSON-LD context (served at /context/v1.jsonld) - website/public/schema/fact.schema.json — JSON Schema for Fact (21 fields, full descriptions) Package: @kndl/memory bumped to 2.0.0 README: full rewrite — mental model, fact shape, install, MCP tools table, storage backends, migration, repo layout, eval quality bar Devils_Advocate.md: v2 retro section — phase-by-phase verdict, resolved/partial/open weak point audit, one-line conclusion
…lish
MCP server / build:
- server.ts: restore .js extensions on @modelcontextprotocol/sdk imports
(linter stripped them; Node.js ESM requires exact extensions for externals)
- stores/index.ts: static import FsFactStore + SqliteFactStore so tsup
bundles them inline — fixes ERR_MODULE_NOT_FOUND from hashed chunk names
(duckdb/supabase remain lazily required, they are optional packages)
Eval runner:
- Fix flag parser: indexOf(-1)+1=0 read args[0] as scenarioFilter="--out",
causing 0 scenarios to match and verdict FAIL with 0 questions evaluated
- Default model: claude-opus-4-5-20251001 (nonexistent) → claude-sonnet-4-6
AI evals example:
- Replace GPT-5/OpenAI facts with Claude 4 family facts
(Sonnet 4.5, Sonnet 4.6 1M, Opus 4.7 — supersession + regression + derivedFrom)
Website:
- ExplorerPage: force-directed graph view; Cards/Graph toggle; drag nodes;
click to show node detail panel (mobile: slides up from bottom)
- ExplorerPage: mobile optimisation — toolbar stacks, domainSelect stretches,
node detail becomes bottom sheet on ≤680px, stats hidden on ≤420px
- ExamplesPage: whole card is now a Link (not just "Explore →" button)
- LandingPage: use case cards redesigned with SVG icons + tag badges + "Explore in graph →"
- LandingPage: CTA cards redesigned with SVG icons, colour-coded by section,
richer description line, "Go →" explicit action
- McpPage: corrected tool list (11 not 9), correct storage backends
(fs/sqlite/duckdb/supabase — no postgres), correct HTTP port (8000 not 7654),
correct install path, correct remote sync CLI syntax
- SkillPage: numbered 2-step install section (CLI first, then skill bundle);
SKILL.md fixes applied (KNDL_STORAGE env var, correct build commands, npx removed)
- EvalPage: runtime fetch of /eval/results.json instead of static import;
loading spinner + error card; threshold marker on pass bar;
NOT_RUN instructions now say "run the runner, redeploy" — no code editing needed
Makefile: kndl-eval writes to website/public/eval/results.json; publish-eval target
SKILL.md: correct setup (pnpm build + npm link), KNDL_STORAGE, bare kndl commands
Nav (all pages):
- Hamburger menu on ≤680px: 7 links collapse behind animated burger button
- Full-width drawer with active-link highlight; closes on tap
- Desktop layout unchanged
Mobile CSS (all pages, ≤680px breakpoint):
- ProtocolPage: container padding 16px, pre font 11px, table cells tighter
- McpPage: storageTable collapses to 1-col, storageWhen badge hidden,
archGrid 1-col, toolRow stacks, layerPills wrap
- EvalPage: summaryGrid stacks vertically, qArchetype column hidden on mobile,
notRunCard padding reduced
- SkillPage: installStep stacks (number above body), downloadCard stacks,
installCode font 11px
- ExamplesPage: container padding 16px, cardDesc smaller font
- LandingPage: heroLogo 36px on 420px, ctaRow 1-col, factCode overflow-x auto
- ExplorerPage: existing fixes kept (toolbar, nodeDetail bottom sheet)
OG / social meta:
- index.html: all v1 text replaced with v2 ("The Format Anthropic Memory Was
Waiting For"); og:image:height corrected 864→860; og:image:alt + twitter:image:alt
added; removed dead links (spec/SPECIFICATION.md, kndl.ebnf, llms-full.txt);
JSON-LD updated — SoftwareSourceCode→SoftwareApplication, TechArticle for
deleted /spec/full removed; noscript updated to v2 three-layer mental model
- SEO.tsx: softwareSourceCodeSchema updated to SoftwareApplication with v2 desc
- kndl.png: confirmed 1860×860, served correctly for both root and website
README.md:
- Full rewrite: centered banner image + badges at top; hook paragraph before
any technical detail; "Get started in 60 seconds" as first section;
Why-not-markdown table with aligned columns; complete CLI workflow example;
HTTP server + Goose/LM Studio multi-agent configs; repo layout tree
- Appendix A: full strategic analysis (pre-pivot devil's advocate) with v2 outcome annotations on each weak point, competitor table, alternative paths, and use-case domain ranking - Appendix B: v2 retro — what the plan said vs. what shipped, resolved/ open/partial breakdown, one-line verdict - §13 cross-reference updated: Devils_Advocate.md → Appendix A - Devils_Advocate.md deleted (content lives in v2.md)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Full pivot to v2