Memory Engineering

July 20, 2026 · View on GitHub

Stop re-explaining the repo. Design the memory. Get a score.

Memory engineering is the practical discipline of giving AI coding agents durable, trustworthy, budgeted recall across sessions — without poisoning context or burning tokens.

Companion to:

Layer	Repo	Unit
Context	essays / runtime	One inference window
Harness	harness-foundry	One agent run
Memory	this repo	What persists between runs
Loop	loop-engineering	What keeps prompting over time
Fleet	fleet-engineering	Populations of agents

# 5 minutes to Memory Ready
npx @cobusgreyling/memory-init .
npx @cobusgreyling/memory-audit . --suggest

Until packages are published, run from a clone:

git clone https://github.com/cobusgreyling/memory-engineering.git
cd memory-engineering
node tools/memory-init/cli.js /path/to/your-project --pattern session-scratchpad
node tools/memory-audit/cli.js /path/to/your-project --suggest

Start here (pick your pain)

Symptom	Start with
Agent forgets decisions overnight	Session Scratchpad
“We already decided X” fights	Project Episodic Log
Wrong facts keep getting reloaded	Durable Facts Store
Context window bloated by history	Retrieval Budget
Stale / poisoned memory	Memory Hygiene Loop
Already running loops	docs/with-loop.md

Unsure? Pattern picker · Quickstart

Why this matters

Loop engineering designs the system that prompts. Memory engineering designs what that system is allowed to remember.

Without it:

Every morning is a cold start
Agents re-litigate settled decisions
Long STATE.md files become untrusted sludge
Token bills climb while quality falls

With it:

Explicit memory tiers (scratch → episodic → durable)
Write / recall / hygiene skills with verification
A Memory Ready score (M0–M3) you can CI-gate
Same house style as loop/fleet: patterns, starters, init + audit

The four memory tiers

Tier	Lifetime	Trust	Example
Scratch	This session	Low	Working notes, open questions
Episodic	Days–weeks	Medium	What happened, decisions, handoffs
Durable facts	Until revoked	High	Stack, owners, invariants, “never do X”
Retrieved	Per inference	Variable	Chunks pulled under a budget

Rules of thumb:

Scratch is cheap to write, expensive to promote.
Durable facts need a human or verifier gate.
Retrieval without a budget is just context spam.
Hygiene is a loop — memory rots.

Getting started (5 minutes)

# 1. Scaffold memory spine + skills
node tools/memory-init/cli.js . --pattern session-scratchpad --tool grok

# 2. Score readiness
node tools/memory-audit/cli.js . --suggest

# 3. Week one: write scratch + episodic only — no auto-promotion to durable

Artifacts created:

File	Role
`MEMORY.md`	Posture: tiers, write policy, hygiene cadence
`MEMORY-STATE.md`	Live catalog of what is remembered
`memory-budget.md`	Token / entry caps
`memory-constraints.md`	What must never be stored
`memory-run-log.md`	Append-only ops log
`skills/`	recall / write / hygiene / verifier skill stubs

CLI surface

Command	Job
`memory-init`	Scaffold pattern + templates + skills
`memory-audit`	Memory Ready score + suggestions
(planned) `memory-compact`	Summarize episodic under budget
(planned) `memory-promote`	Propose scratch → durable with verifier gate
(planned) `memory-mcp-server`	MCP tools: recall / write / list / hygiene

Full flags: docs/cli-surface.md

Memory Ready (preview)

Memory Ready: 54/100  ·  M1 structured

OK   MEMORY.md present
OK   MEMORY-STATE.md present
OK   write + recall skills
WARN no hygiene skill
WARN no memory-budget.md
→ node tools/memory-init/cli.js . --pattern memory-hygiene-loop

See docs/memory-ready-score.md.

Stack diagram

Model  →  Context pack  →  Memory tiers  →  Loop  →  Fleet
          (this turn)      (across turns)   (time)   (population)

Contributing

See CONTRIBUTING.md. Pattern PRs welcome — use templates/pattern-template.md and register in patterns/registry.yaml.

Sources

resources/sources.md

License

MIT — LICENSE