# Phase 5: Hindsight Memory Provider - Discussion Log

> **Audit trail only.** Do not use as input to planning, research, or execution agents.
> Decisions are captured in CONTEXT.md — this log preserves the alternatives considered.

**Date:** 2026-06-14
**Phase:** 5-Hindsight Memory Provider
**Areas discussed:** Hindsight mode, Memory integration, Migration strategy, Recall behavior, Retain frequency

---

## Hindsight Mode

| Option | Description | Selected |
|--------|-------------|----------|
| Cloud | Send data to Hindsight Cloud API. Need account at ui.hindsight.vectorize.io. Zero local infra but external dependency + API cost. | |
| Local Embedded | Hermes spins up local PostgreSQL daemon. Uses existing OpenRouter key for LLM extraction. No external data send. ~200MB download, daemon idles after 5 min. | ✓ |

**User's choice:** Local Embedded
**Notes:** Privacy preference — no external data send. Existing OpenRouter key reused.

---

## Memory Integration

| Option | Description | Selected |
|--------|-------------|----------|
| Hybrid | Auto-inject relevant memories before each turn + expose 3 hindsight tools (retain, recall, reflect) to agent | ✓ |
| Context | Auto-injection only, no tools exposed | |
| Tools | Tools only, no auto-injection | |

**User's choice:** Default (Hybrid)
**Notes:** Wants agent to have both automatic context and on-demand search capability.

---

## Migration Strategy

| Option | Description | Selected |
|--------|-------------|----------|
| No migration | Built-in and hindsight run in parallel. No data sync. Hindsight builds fresh knowledge graph. | ✓ |
| One-shot seed | Manual migration script to port MEMORY.md contents into hindsight | |
| Keep both fallback | Built-in as write-only fallback for critical facts | |

**User's choice:** No migration needed
**Notes:** Accepts that hindsight starts fresh. Built-in memory continues working in parallel.

---

## Recall Behavior

**User's choice:** Agent discretion — with latency constraint
**Agent's recommendation:** `recall_budget: low`, `recall_prefetch_method: recall` (raw fact search, no LLM synthesis), `auto_recall: true`, `recall_types: observation` (default)
**Notes:** User wants responsive memory recall that doesn't add noticeable latency. Agent recommended fast-path settings accepted.

---

## Retain Frequency

**User's choice:** Agent discretion — with latency constraint
**Agent's recommendation:** `retain_async: true` (critical for latency), `retain_every_n_turns: 5` (80% overhead reduction), `auto_retain: true`
**Notes:** User wants memory extraction without blocking the agent loop. Agent recommended async + batched retain settings accepted.

---

## the agent's Discretion

- Bank configuration (bank_id, bank_mission, bank_retain_mission) — use defaults, planner can recommend tuning
- Daemon startup and monitoring — standard Hermes daemon management

## Deferred Ideas

None.