The re-explanation tax is real

You explain architecture on Monday. By Tuesday, your assistant forgot half of it. By Friday, you’re writing the same primer again.

Why this happens

Coding assistants optimize for current window + short conversation history. They don’t preserve durable project context unless you build that layer.

What to store

repository topology
ownership map
non-obvious constraints
ADR summaries
coding conventions

Example memory entries

- payments-service uses outbox pattern; never publish events inside request transaction
- all external API calls go through /lib/http/client.ts wrapper
- tenancy boundary is org_id, never workspace_id

Retrieval pattern for coding tasks

When prompt includes file paths, architecture terms, or migration words, fetch:

relevant ADR memories
prior bug postmortems on same subsystem
team conventions for touched files

Why vector-only fails

Similarity alone may return “sort of related” notes. Add metadata filters (repo, subsystem, env) and graph edges (dependsOn, ownedBy).

Practical OpenClaw flow

wake briefing injects current project context
on implement/fix/refactor intents, run targeted recall mode
cap memory injection to avoid prompt bloat

Developer UX win

You stop writing this in every session:

“Remember, we use event sourcing in billing and CQRS read models in analytics.”

The system already knows.

Team impact

New engineers ramp faster because the assistant carries institutional memory that usually lives in fragmented docs and tribal knowledge.

Short checklist

define memory schema per repo
add checkpoint on major merges
run recall probes weekly

Do this and your coding assistant starts behaving like a teammate with continuity.