Bank presets (memory starter kits)

When you create or update a bank, you can pass bank_preset on the same request body as other CreateBankRequest fields (OpenAPI: PUT / PATCH /v1/default/banks/{bank_id}; see also Memory banks).

Presets are curated defaults that:

1. Merge into bank config before any explicit fields you send (explicit fields override the preset).
2. Optionally seed operator-facing content (pinned mental models and one directive) for specific presets.

They exist so teams do not have to hand-copy retain/reflect/observation tuning for common workflows (for example repository + Codebases).

Supported values

bank_preset	Behavior
(unset)	No preset merge; only fields you send apply.
codebase	Applies repository-oriented retain/reflect/observation settings and idempotently seeds developer guides.

Unknown strings are ignored (no error), so clients can safely forward user-selected labels.

codebase preset

Configuration merged

The preset sets (unless overridden by your request):

• retain_extraction_mode → custom, with retain_custom_instructions focused on technical facts from code chunks (paths, symbols, contracts, risk).
• retain_mission, reflect_mission, observations_mission — short missions aligned with evidence-backed engineering memory.
• enable_observations → true so cross-file themes can consolidate after retain.

Implementation lives in atulya_api/bank_presets.py (BANK_PRESET_CONFIG).

Seeded content (idempotent)

On PUT or PATCH when bank_preset is codebase, the API runs seed logic once per logical artifact:

• One directive — evidence-first answering (path/line provenance; no inventing structure when memory is thin). Skipped if a directive with the same name already exists for the bank.
• Three pinned mental models — short Markdown guides: review → memory workflow, how to read code-backed memories, operating tips. Each uses a stable id per guide; if that id already exists for the bank, the insert is skipped (concurrent-safe).

Tags include preset:codebase so you can filter or delete starter content later.

Seeding failures are non-fatal (bank creation still succeeds); check server logs if guides are missing.

Control Plane

On the Dashboard quick-start flow, Memory starter kit lets you choose Code repository (sends bank_preset: "codebase" on bank create). The Control Plane POST /api/banks forwards the JSON body to the API (including bank_preset).

API example

curl -X PUT "http://localhost:8000/v1/default/banks/my-codebase-bank" \
  -H "Content-Type: application/json" \
  -d '{"bank_preset": "codebase"}'

Combine with explicit overrides in the same body; explicit keys win over the preset.

When not to use a preset

• You already maintain bank config in code or Terraform and want full control — omit bank_preset and set fields explicitly.
• You need a minimal bank for tests — use default creation without a preset.

---

Roadmap (future improvements)

This area is intentionally small and data-driven. Likely next steps:

• More presets — e.g. docs-first, incident response, or “customer CRM” kits with different retain/reflect defaults.
• Preset versioning — pin bank_preset to a version string when behavior changes so existing banks stay reproducible.
• Control Plane editing — pick a preset, then tweak fields in UI before save, with a diff preview against the preset base.
• CLI parity — atulya bank create flags mirroring bank_preset.
• Optional LLM refresh — after the first successful codebase import, optionally refresh a guide mental model via reflect (today seeds are static Markdown + embeddings only, no async reflect jobs at signup).

If you extend presets in code, keep the single source of truth in bank_presets.py, regenerate OpenAPI + clients, and update this page.