Durable memory
Thin core + domain sets for a fast, stack-aware project start. Cross-agent (Claude Code, Cursor, Codex, Gemini) via the SKILL.md standard. MIT.
npx -y skills add Neznakometz/StackForge --skill durable-memoryAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
L3 durable memory for the agent across sessions — SQLite + agent-memory.md, consolidation and cleanup. Apply when you need persistent project memory (patterns, paths, architectural decisions) beyond the core checkpoints (L2).
SKILL.md
1.8 KB, as published. Nobody here has run it
L3 — durable memory
Extends the core memory (L1 context + L2 checkpoints). L3 doesn't die between sessions.
Storage
- SQLite (
store/agent.db) — the machine-readable source of truth (if an MCP/plugin for memories is available). agent-memory.md— a human-readable mirror for review.- If SQLite is unavailable — work via
agent-memory.mdonly (simpler, less reliable).
What lives in L3
Paths to important files/folders, architectural decisions ("why a webhook, not polling"), style rules, the user's behavioural patterns, the project's long-term goals.
Expiration policy
- Keep permanently: paths, architecture, patterns, style rules.
- Delete after 30 days: one-off findings, temporary solutions, observations without generalization.
- Permanent-storage markers:
permanent/important/keep(in the text or a tag) — protect a record from cleanup.
Rules for writing to agent-memory.md
- Specify a
scope: user / project / team / session. - Replace outdated records on the same topic — don't accumulate contradictions.
- Normalize dates to absolute format (YYYY-MM-DD).
- Compact facts and patterns, not raw chat logs. A task's temporary noise → L2, not here.
Cycle
- Before a task — read
agent-memory.mdfor relevant knowledge. - After — if you learned a permanent pattern/fact, add it (a generalization, not junk).
- Maintenance commands:
/dream(consolidation),/warmup(cache),/memory-gc(cleanup).
Record format
### YYYY-MM-DD — topic_key — [PERMANENT]
- Scope: project
- Content: [a specific fact or pattern]