---
id: CYA-WP-0003
type: workplan
title: "Contextual Memory Activation and Retrospection Loops for Continuous Optimization"
domain: agents
repo: can-you-assist
status: done
owner: grok
topic_slug: foerster-capabilities
created: "2026-05-27"
updated: "2026-05-27"
state_hub_workstream_id: "ac9a9d42-4db9-44d0-9b41-5e16d87c65c8"
---

# CYA-WP-0003: Contextual Memory Activation and Retrospection Loops for Continuous Optimization

## Goal

Build on the real memory foundation delivered in CYA-WP-0002 to make memory **actively useful** for context-aware assistance, with two primary advances:

1. **Directory- and project-bound memory activation**: Automatically and intelligently surface relevant memory when the user is working in a specific directory or project, without requiring explicit recall every time.
2. **Regular retrospection loops**: Enable structured, recurring reflection sessions between the user and `cya` that turn memory into a vehicle for continuous improvement — reviewing what worked, setting interaction goals, and evolving preferences over time.

This workplan directly addresses two remaining gaps identified in the post-0002 Intent-vs-Scope analysis:
- Moving from passive memory storage to **proactive, contextually activated** memory.
- Establishing the foundation for true **longitudinal adaptation** and a user-driven optimization loop.

## Background & References

- `INTENT.md` — especially the "Personalized Console Helper" section and the explicit call for `phase-memory` to support "project-specific memory" and "recurring workflows".
- `MemoryVision.md` — defines primary memory kinds relevant to cya, including Project/Directory Memory and Workflow Recipes.
- `history/2026-05-27-CYA-Intent-Scope-Gap-Analysis-Post-0002.md` — identifies remaining gaps in contextual activation and longitudinal value after 0002.
- CYA-WP-0002 — delivered the real (persisting) memory implementation and explicit ports.
- Current memory ports in `src/cya/memory/__init__.py` (the `scope` parameter is the natural starting point for directory binding).
- phase-memory concepts (phases, profiles, activation planning, dry-run-first).

## Non-Goals (for this slice)

- Full autonomous agentic behavior or background memory harvesting.
- Deep semantic understanding of code or notes (still out of scope).
- Building retrospection as a fully automated scheduled process (user-initiated or explicitly prompted is preferred for control).
- Replacing the explicit port seam with direct phase-memory calls (we continue to evolve through the ports).
- Rich multi-turn conversational state beyond what scoped memory + retrospection can provide.

## Task Breakdown

### T01 — Refine the conceptual model for contextual activation and retrospection loops

```task
id: CYA-WP-0003-T01
status: done
priority: high
state_hub_task_id: "335595b4-96fe-401f-ab4c-0824a6ba3f05"
started: "2026-05-27 ralph iter 1"
completed: "2026-05-27"
```

**Done** — produced `docs/cya-memory-activation-and-retrospection-concept.md`.

- Deeply aligned the two ideas with INTENT.md (Personalized Console Helper, project-specific memory, recurring workflows, user control) and MemoryVision.md (Project/Directory Memory, Workflow Recipes, activation planning, explainability requirements).
- Defined clear models:
  - **Memory Activation**: Automatic, scope-aware (cwd + git root) surfacing of relevant items (preferences, patterns, safety signals) with strong provenance and user control.
  - **Retrospection Loops**: Deliberate reflection sessions that produce higher-order memory (`interaction_goal`, retrospection outcomes) which preferentially influence future activation and behavior.
- The two mechanisms reinforce each other: normal use generates candidates for retrospection; retrospection produces high-quality memory that improves future activation.
- Identified minimal port/data model extensions for T02 (richer `kind` support and activation hints in recall, standardized retrospection record kinds).
- Non-negotiables (explainability + safety invariants) explicitly preserved.

**Acceptance criteria met**:
- Clear, written concept document exists and is consistent with INTENT and MemoryVision.
- The ideas are shown to reinforce each other in a user-controlled continuous optimization loop.

### T02 — Extend memory ports and data model for activation and retrospection

```task
id: CYA-WP-0003-T02
status: done
priority: high
state_hub_task_id: "6f50bdf4-6252-4b93-9697-407ef432cd90"
started: "2026-05-27 ralph continuation (after T01 review)"
completed: "2026-05-27"
```

**Done** — implemented in `src/cya/memory/__init__.py`.

- Added `kind` parameter to `remember_preference` (defaults to "preference" for full backward compat).
- Added `KIND_*` constants and `remember_retrospection_outcome()` convenience helper.
- Enhanced `recall_preferences` with better `kinds` filtering + new `activation_context` parameter for smarter directory/project-aware activation.
- Improved `export_memory` with optional `kinds` filter and `by_kind` summary.
- All changes maintain 100% backward compatibility with existing call sites.
- Verified manually (roundtrips for preferences, retrospection records, kind filtering, and activation hints all work).

**Acceptance criteria met**:
- Ports and data model now support the T01 concepts.
- Everything remains fully user-visible/editable (still plain JSON under `~/.config/cya/memory/`).
- No breaking changes.

### T03 — Implement directory- and project-bound memory activation

```task
id: CYA-WP-0003-T03
status: done
priority: high
state_hub_task_id: "45731b48-74a5-485b-bd56-72f387db3846"
started: "2026-05-27 ralph continuation (after T02)"
completed: "2026-05-27"
```

**Done** — implemented in `src/cya/orchestrator.py`.

- Memory recall now passes `activation_context` containing cwd + git_root (when available from the envelope).
- The T02-enhanced `recall_preferences` uses this for smarter directory/project-bound activation (boosting matching scopes).
- Improved `--explain-context` panel now shows activation context and sample activated keys.
- Final output line simplified and updated to reflect activation.
- Users can influence via normal `remember_preference(..., scope=...)` and the existing export/forget tools (more UX in T04/T07).

**Acceptance criteria met** (for this slice):
- Memory is now automatically activated based on working directory/project.
- Activation is visible (in explain panel + provenance) and controllable via existing memory tools.

### T04 — Build the retrospection interaction flow

```task
id: CYA-WP-0003-T04
status: done
priority: high
state_hub_task_id: "fb63edc4-1e3f-4964-a2b6-19b29d00ffd8"
started: "2026-05-27 ralph continuation (after T03)"
completed: "2026-05-27"
```

**Done** — implemented.

- Added `cya retrospect` subcommand in `src/cya/cli/main.py`.
- Implemented `run_retrospection()` in `src/cya/orchestrator.py`:
  - Reviews recent memory in the given scope.
  - Guides the user through reflection questions.
  - Records outcomes using the T02 `remember_retrospection_outcome` helper (stored as retrospection / interaction_goal kinds).
  - Uses rich panels for a pleasant terminal experience.
- The command is discoverable (`cya --help` shows the `retrospect` subcommand).
- Outcomes are stored in the same user-controlled memory store and will be activated in future normal sessions (per T03 activation logic).

**Acceptance criteria met** (MVP):
- A user can run `cya retrospect` and capture goals/preferences that affect future assistance.
- The flow is natural, guided, and fully respects user control (they decide what to record).

### T05 — Tests, observability, and graceful degradation

```task
id: CYA-WP-0003-T05
status: done
priority: medium
state_hub_task_id: "f17a4f42-1630-4244-bdd4-c8d732e8de9b"
started: "2026-05-27 ralph continuation (after T04)"
completed: "2026-05-27"
```

- Added comprehensive tests in `tests/test_memory.py` for:
  - Activation logic with `activation_context` (T03)
  - Retrospection outcomes and kind-specific recall/export (T04)
  - Observability (provenance, `by_kind`, activation_context recording)
  - Graceful degradation on bad context or storage issues
- All tests are hermetic (using the existing `isolated_memory` fixture) and introduce **no new external dependencies**.
- Full suite passes cleanly: `pytest tests/test_memory.py -q`

**Acceptance criteria met**:
- Strong test coverage for the new activation + retrospection behaviors.
- `pytest` remains clean with no new external dependencies.

### T06 — Tests, observability, and graceful degradation (completed in T05)

```task
id: CYA-WP-0003-T06
status: done
priority: high
state_hub_task_id: "c7da1a0d-647f-48d9-aba3-d2e62791e05f"
started: "2026-05-27"
completed: "2026-05-27"
```

Tests were completed as part of the previous step (strong coverage for activation, retrospection, observability, and graceful degradation; all tests pass cleanly).

### T07 — Documentation, examples, and handoff

```task
id: CYA-WP-0003-T07
status: done
priority: medium
state_hub_task_id: "d120efff-eaf6-4630-a3d3-3e70a3db4e09"
started: "2026-05-27 ralph continuation (final tasks)"
completed: "2026-05-27"
```

**Done.**

- Heavily expanded the Memory section in README.md with clear before/after examples for:
  - Automatic directory/project-bound activation
  - `cya retrospect` guided sessions and goal capture
- Updated AGENTS.md "Commands" section with the new `cya retrospect` command and 0003 workplan reference.
- Added a substantial "What CYA-WP-0003 Delivered" section to MemoryVision.md documenting the new capabilities and their alignment with the original vision.
- Registered technical debt (retrospection/activation still on local JSON; deeper phase-memory integration is future work).

**Acceptance criteria met**:
- A reader of the README can now understand and try the new memory features (activation + retrospection) with concrete examples.
- The concepts are clearly connected back to INTENT.md and MemoryVision.md.
- Extension points and debt are documented.

## Dependencies & Cross-Repo Coordination

- **phase-memory**: The concepts here are designed to be consumable and eventually enriched by richer phase-memory capabilities (profiles, phases, activation planning). Close coordination recommended during T01/T02.
- **markitect-tool**: May become relevant if retrospection outcomes or directory contexts are expressed as memory profiles.
- State Hub: For tracking this as the direct follow-on to CYA-WP-0002.

## Completion

**Status: done** — all 7 tasks completed via ralph loop. The workplan delivered contextual memory activation (T03), a user-driven retrospection flow (T04), supporting port extensions (T02), tests (T05), and full documentation (T07).

This workplan successfully implemented the two main ideas requested: directory/project-bound memory activation and regular retrospection loops for continuous optimization — directly realizing deeper parts of INTENT.md and MemoryVision.md.

**Status note**: CYA-WP-0003 is complete. The explicit memory seam is now significantly more powerful while remaining fully user-controlled and transparent.