forkmind-archivist

ForkMind Archivist — offload context as an encrypted DAG

Industry context managers treat a full window as a cache-eviction problem: truncate and lose it. ForkMind capsules invert that — persist first, verify, then compact. Nothing is dropped until it is provably on disk.

Core loop

Step	Tool	Rule
1. Save	forkmind_context_save	Pass items + a digest YOU write (≤5 lines, decision-relevant)
2. Verify	forkmind_context_digest	Re-reads from disk. Save is not real until this succeeds
3. Compact	(your window)	Replace archived material with one handle line
4. Probe	forkmind_context_digest	Later: check the digest before restoring anything
5. Restore	forkmind_context_restore	Full, or segmentIds for just one piece
6. Forget	forkmind_context_forget	Irreversible crypto-shred; explicit user ask + confirm: <id> only
7. Move	forkmind_context_export / _import	Portable passphrase-encrypted bundle for another machine/project

Handle line to keep in the window after compacting:

[capsule 9f3ac21b7e04 · "auth-refactor context" · restore via forkmind_context_restore]

When to offload

• Window pressure: long session, old turns no longer active.
• Finished sub-topics: debugging concluded, decision recorded.
• Bulk payloads: file dumps, test logs, tool transcripts already acted on.
• User privacy: sensitive material that should leave the window — save it without a digest (private capsule; manifest carries structure only).

Do NOT offload: the current task's active instructions, unresolved questions, or anything needed in the next few turns.

Guarantees (why this is safe)

• Immutable + acyclic by construction — segments are content-addressed (hash over content + parents, Git-style), so capsules cannot be edited or contain cycles; forkmind context verify <id> re-proves it.
• No plaintext at rest — AES-256-GCM per-capsule keys; the master key lives outside .forkmind/ (~/.forkmind-keys/), so a committed .forkmind/ leaks only ciphertext.
• Crypto-shred delete — forgetting destroys the key first; backups of the ciphertext stay unreadable. Tombstones stop id resurrection.
• Local-first — no cloud, no account, no telemetry; the underlying model and provider are never touched (this all happens in what the client sends).
• RAID redundancy — if forkmind context replicas add <path> targets are configured, capsules mirror automatically (ciphertext only, never keys) and restore self-heals from replicas on primary loss/corruption. Check health with forkmind_context_replicas; after a forget with unreachable replicas, run a sync once they return.