Build Log

Context Vault — Research & Architecture

August 15, 2026

Designed a provider-neutral, user-owned AI memory architecture — research review across 7 systems, a Council pressure-test, and four grounded design docs — with zero code and zero personal data touched.

ExperimentResearchArchitectureAI Memory

systems researched

docs written

council personas run (full depth)

lines of ingestion/product code written

Timeline

Start

Zeb delivered the Charge

A 43-section Charge for a user-owned, provider-neutral AI contextual memory layer, bootstrapped from a real ChatGPT export

Scope

Three scoping questions asked and answered

New experiments/context-vault/ folder; research+architecture only this session; export not yet in hand

Crucible

Council review — architect, engineer, qe, full depth

One batched council call, full raw output printed verbatim, synthesized into 5 concrete design changes

Research

3 parallel agents against primary sources

Letta/MemGPT + Mem0; MemX + LongMemEval + LoCoMo/-Plus; Solid + MCP + temporal/graph research

Design

4 docs written, grounded in both the Council and the research

research.md, architecture.md, memory-model.md, threat-model.md, evaluation-plan.md

What shipped

experiments/context-vault/ — new standalone folder, not wired into any ForgeKit product

experiments/context-vault/data/.gitignore + README.md — personal data exclusion in place before any real export exists

experiments/context-vault/docs/research.md — prior-art review: Letta/MemGPT, Mem0, MemX, LongMemEval, LoCoMo/LoCoMo-Plus, Solid, MCP, plus temporal/graph memory research (Graphiti, arXiv 2606.01435)

experiments/context-vault/docs/architecture.md — ADR: SQLite+FTS5+optional-embeddings, a three-table (not two) candidate/canonical split, retrieval pipeline shape (fuse-then-reject), MCP exposure (Resources+Tools split by interaction shape), what's deliberately not built

experiments/context-vault/docs/memory-model.md — schema: 5 canonical primitives (not 8 storage shapes), provenance as a separate many-to-many table, dual timestamp pairs (valid-time vs ingestion-time), explicit-vs-inferred evidence weighting, forgetting vs correction as distinct operations

experiments/context-vault/docs/threat-model.md — concrete data-flow answers (what leaves the machine, when, logged how) rather than a generic risk-category list

experiments/context-vault/docs/evaluation-plan.md — ~50-question eval design with a new knowledge-update category (LongMemEval), matched-pair abstention construction, and the two-separate-metrics fix for abstention-gaming (correct-abstention-rate + false-abstention-rate, never blended)

experiments/context-vault/README.md — orientation doc, non-goals, what's next

“The mechanism is worth borrowing; the marketing numbers are not evidence for adopting it.”
— Research agent, on Mem0's disputed LOCOMO benchmark claims — the discipline this whole research pass tried to hold to throughout