> The user’s working memory must belong to the user — not to any assistant interface, vendor, or opaque system.
`cya` must never become a memory silo. All memory, history, preferences, and adaptation behavior must flow through explicit, inspectable ports provided by `phase-memory`.
## Relationship to phase-memory
`phase-memory` is the **profile-driven memory operating layer** for the broader ecosystem. It is responsible for:
- Interpreting **Markitect memory profiles** as executable runtime plans.
- Modeling memory in distinct **phases**:
-`ephemeral` — short-lived, high-turnover (e.g., current conversation window)
-`fluid` — recent but still malleable context
-`stabilized` — reviewed and relatively durable knowledge/preferences
- Planning and executing (or dry-running) **lifecycle actions**:
- Retention
- Refresh
- Compaction
- Stabilization
- Activation (under token/item budgets)
- Policy, audit, and review gates around memory changes.
`cya` is a **consumer** of this layer. It does **not** own memory semantics, storage, or lifecycle policy.
## How cya Will Use Memory
`cya` will treat memory as a first-class, explainable input to every assistance cycle, while keeping ownership and control firmly with the user (via phase-memory).
- **Source Provenance** — metadata about where knowledge came from (files, git history, previous sessions)
### Integration Model (Building on T05 Ports)
The four explicit ports defined in T05 (`cya/memory/__init__.py`) are the permanent integration surface:
-`remember_preference(key, value, scope)`
-`recall_preferences(scope, task_class)`
-`forget(scope, keys)`
-`export_memory(scope)`
Over time these will be wired to real `phase-memory` capabilities. `cya` will additionally:
1.**Produce structured memory events** (conversation turns, accepted suggestions, explicit user feedback, safety decisions) that phase-memory can ingest as fluid memory.
2.**Request context packages** via phase-memory’s activation planning (respecting user-defined profiles and token budgets).
3.**Surface memory influence** in responses — users should be able to see “this suggestion was shaped by your saved preference for concise git output + recent project conventions.”
4.**Honor user-initiated lifecycle actions** (export, reset, compact, stabilize) exposed through the ports or dedicated `cya memory ...` subcommands in the future.
### Profile-Driven Behavior
In the longer term, users (or the system) will be able to supply or generate **Markitect-compatible memory profiles** that tell phase-memory how `cya` should behave regarding their memory. Examples of profile intent that matter to cya:
- “Keep my last 30 days of shell interactions fluid, but stabilize any workflow I have used more than 5 times.”
- “Treat everything in `~/notes/projects/` as stabilized by default, but require review before anything becomes rigid.”
- “Aggressively compact conversational memory after 48 hours unless I mark it important.”
- “Never let total activated context for a single `cya` request exceed 8k tokens without explicit approval.”
`cya` will request activation plans and context packages that respect these profiles.
## Safety and Explainability Requirements
Because memory directly influences suggestions:
- Every response that used non-ephemeral memory must be able to explain **which** memory items influenced it and why.
- Memory-driven suggestions must still pass through the T03 risk classifier.
- Users must be able to temporarily or permanently disable memory influence for a session or scope (`cya --no-memory ...` or equivalent).
- Export and inspection of memory must be first-class and low-friction.
## Current State (Post MVP)
As of the end of CYA-WP-0001:
- Only the four thin no-op ports exist (T05).
- No actual memory is persisted or recalled by `cya`.
- The orchestrator and safety layer are ready to consume memory signals when they become available.
- The architecture deliberately avoids any hidden or local-JSON fallback memory (per explicit operator direction).
This is the correct starting point. Real integration will be done as a subsequent slice, using the ports as the stable contract.
## Open Questions & Future Work
- How will `cya` generate or help users author memory profiles?
- What is the right granularity for “project memory” vs global memory?
- How should safety decisions and user overrides themselves become first-class memory citizens?
- Can/should `cya` participate in compaction and stabilization planning, or should it only consume the results?
- What observability (cost, token usage, memory influence) should be exposed at the CLI level?
## Success Criteria (Longer Term)
`cya` + `phase-memory` together are successful when a user can say:
- “This assistant actually knows my preferences and project conventions, but I can see exactly what it knows and change it whenever I want.”
- “I switched LLM providers and my assistant still feels like *mine*.”
- “I asked it to forget everything about project X last month, and it actually did.”
Building directly on the 0002 integration contract (below), the 0003 contextual activation + `cya retrospect` continuous optimization loop, and the post-0004 gap analysis (memory moved from large gap to small-medium), deep research into agentic/self-improving memory architectures was performed and persisted.
The research synthesizes three canonical patterns for closed self-improving loops in LLM agents (Trial → Evaluation/Feedback → Reflection/Synthesis → Memory Update → Improved future behavior):
- **Profile 1 — Reflexion-style verbal self-improvement** (lightweight, high-explainability): Store natural-language self-reflections/lessons (leveraging existing `KIND_RETROSPECTION` + `remember_retrospection_outcome` + `cya retrospect`). Preferentially activate them via kinds + activation_context in future turns. Plain-English "verbal reinforcement" that users can inspect and edit.
- **Profile 2 — Generative-Agents-style hierarchical synthesis**: Treat assistance outcomes, retrospections, and explicit remembers as an episodic stream. Periodic (user-triggered or planner-driven) LLM synthesis produces higher-order abstractions (project conventions, workflow patterns, "in this scope we always...") with citations, stored in stabilized phases. Multi-factor retrieval (recency + importance + relevance + profile/scope match) + the existing activation boost.
- **Profile 3 — Procedural / meta-policy evolution** (highest leverage, highest guardrail needs): First-class evolvable "how I should behave" rules and procedures as a distinct tier. Meta-reflection (after retrospection or explicit "improve my rules") proposes patches to the procedural layer. Strong dry-run + user veto + safety impact analysis required; changes can only tighten (or maintain) the rule-based risk posture by default.
Detailed definitions, cya-specific implementation mappings (ports, kinds, orchestrator wiring, safety invariants), capability matrix, and phase-memory interface requirements live in the persisted research document and are executed via CYA-WP-0005 tasks T02–T05.
This directly addresses several Open Questions above (profile authoring, granularity of project memory, safety signals as first-class memory, observability of memory influence, participation in planning/compaction) while preserving all cya invariants (user control, full provenance/explainability, rule-based safety that memory can only strengthen).
**Next in this workplan:** T02 formalizes the current post-0003 implementation as explicit **Profile 0** baseline (the stable foundation everything else builds on). T03 adds the full profile definitions + matrix. T04 delivers the concrete optimization suggestions for the phase-memory sister repo.
See also the full research artifact and CYA-WP-0005 for acceptance criteria and cross-links.
## Profile 0 Baseline (Post-0003 / Current Shipped)
**Status:** Shipped and stable as the production memory implementation (CYA-WP-0002 T02 real JSON + CYA-WP-0003 contextual activation + retrospection extensions). This is the explicit foundation on which Profiles 1–3 will be built.
**Guiding Principle (Profile 0):** Deliver real, user-controlled, contextually activated, longitudinally improving memory *today* using a high-quality local approximation, while keeping the integration seam (profile, kinds, activation_context, provenance, phase hints) completely stable and ready for eventual replacement by full phase-memory planners, graph store, and lifecycle rules. No hidden state; everything is user-inspectable and user-owned.
### Backing Store & Persistence
- Location: `~/.config/cya/memory/<scope>.json` (one file per scope, default "cwd").
- Format: Simple list of dict records. Fully human-readable and editable by the user.
- Fields per record (typical): `key`, `value`, `ts` (epoch), `scope`, `profile`, `kind`.
- User can `cat`, edit, or delete these files at any time; `cya` will respect the changes on next run.
- When full phase-memory is wired, this backing will be replaced by the durable graph/event store while preserving the exact same high-level port behavior and return shapes.
### Public API (the stable cya ↔ phase-memory seam)
All entry points live in `src/cya/memory/__init__.py`:
- **Constants** (standard kinds used by cya and `cya retrospect`):
-`KIND_PREFERENCE`
-`KIND_RETROSPECTION`
-`KIND_INTERACTION_GOAL`
- **Core functions** (signatures as of 2026-05, all support `profile` for future multi-profile use):
-`remember_retrospection_outcome(...)` — convenience wrapper that chooses the right kind for higher-order memory from reflection sessions.
**Activation logic (0003):** When `activation_context` (populated by the orchestrator from `ContextEnvelope` with `cwd` + git root) is supplied to `recall_preferences`, items whose `scope` or `profile` matches are boosted to the front of the result list. This makes directory/project-bound memory feel proactive without any user having to explicitly recall it.
**Retrospection as first-class input (0003):** `cya retrospect` (guided flow) produces records with special kinds that receive preferential activation in future turns. These are the seed for the self-improving loop described in the 2026-05 research.
-`phase` hint returned (currently always "fluid" for the local store; will become meaningful once phase-memory lifecycle is wired).
- Graceful degradation: on any error the ports log a loud warning to stderr and return safe empty/default values. Never crash the assistance path.
- All memory influence is visible via `--explain-context`.
### Safety & Explainability Invariants (non-negotiable for Profile 0 and all future profiles)
- Memory signals are **only additive** to caution. They may append rationale or force confirmation in the rule-based `RiskClassifier`, but they never downgrade a risk level or bypass mandatory explicit user confirmation for non-SAFE commands.
- Every memory-influenced response can explain exactly which items were activated, why (activation_context match, kind boost, recency, retrospection provenance), and what phase they came from.
- Users can always opt out per-request (`--no-memory` or equivalent) or globally inspect/forget via the ports + future `cya memory` subcommands.
### Usage Sites in the Current System
-`src/cya/orchestrator.py`: Calls `recall_preferences` with activation_context on every assistance request; injects results into the `ContextEnvelope` passed to the LLM; renders memory influence in `--explain-context` panels.
-`src/cya/cli/main.py` (retrospect subcommand): Uses `recall_preferences` + `remember_retrospection_outcome` to close the user-driven continuous optimization loop.
- Risk classifier: Receives memory context and applies only the "increase caution" rules.
### Relationship to Profiles 1–3 and phase-memory
Profile 0 is deliberately a **complete, usable, production-quality local implementation** of the seam defined in the 0002 T01 contract (refined in 0003). It already delivers the INTENT/SCOPE vision for contextual + longitudinal memory with excellent explainability and zero hidden state.
Future profiles will layer on top:
- Profile 1 adds verbal reflection storage + preferential activation (mostly a small delta on existing retrospection kinds + `cya retrospect`).
- Profile 2 adds episodic capture + synthesis passes (new kinds + calls into phase-memory reflection planners).
All of them must continue to satisfy the invariants above and must continue to work against the exact same port signatures (or compatible extensions).
**See:** `history/2026-05-28-CYA-Agentic-Memory-Research-Variations.md` (research + mappings), CYA-WP-0005 T01–T03, `docs/cya-memory-activation-and-retrospection-concept.md`, current `src/cya/memory/__init__.py` (the reference implementation), and the 0002 integration contract below (the original seam that Profile 0 realizes).
## Profiles 1–3: Definitions and cya Integration Plans
**Date:** 2026-05-28 (CYA-WP-0005 T03)
**Source:** Synthesized from `history/2026-05-28-CYA-Agentic-Memory-Research-Variations.md` (the deep research artifact). These are the authoritative working definitions for the self-improving memory profiles. All three preserve cya's non-negotiable contract: user-controlled, fully explainable, safety-first (memory only increases caution), explicit seam to phase-memory.
The three profiles are ordered by increasing agentic power and implementation cost.
**Status:** Production (CYA-WP-0006). Shipped with guided capture in `cya retrospect`, `cya memory reflections`, explicit compaction, and budget-capped surfacing in responses and `--explain-context`.
**Intent:** Enable lightweight, high-explainability self-improvement by capturing and preferentially activating natural-language "lessons" and verbal reflections from user interactions and retrospection sessions.
**Core Loop (condensed):**
1. Normal assistance with current Profile 0 activation.
2. Outcome capture (accept/revise/reject or explicit `cya retrospect`).
3. Verbal reflection generated (1–3 concise lessons in plain English).
4. Stored with new or extended `kind` (e.g. `reflection` or `verbal_lesson`) via `remember_retrospection_outcome` / new helper.
5. Future recall boosts these kinds + activation_context; they are prepended with high salience.
6. Visible in `--explain-context` and final output ("3 verbal reflections influenced this...").
- Small delta: lightweight "capture lesson" step at end of retrospect (or auto after notable outcomes); new `kind` values; preferential activation logic tweak.
- Orchestrator already has the wiring to surface them.
**Delta Required:** Minor enhancements to retrospect CLI flow and recall boosting; new `kind` constant(s).
**Phase-memory Fit:** Minimal for MVP — good native `kinds` filter + provenance. Later: a lightweight "reflection planner" for compaction of duplicate lessons.
**Safety / Explainability:** Reflections that touch risky patterns still go through the rule-based RiskClassifier. All items carry full provenance. Users can inspect/edit/forget them directly.
**See also:** Research doc Variation 1, Shinn et al. (Reflexion), current `docs/cya-memory-activation-and-retrospection-concept.md`.
**Intent:** Move from passive storage to active synthesis of higher-order knowledge (project conventions, recurring workflows, "in this scope we always...") so the assistant becomes meaningfully smarter over time with citations back to source events.
**Core Loop:**
1. Episodic capture of every turn/outcome/retrospection (structured records with ts, kind, scope, activation_context, payload, provenance).
2. Periodic synthesis (user-triggered via `cya retrospect --synthesize`, or phase-memory planner): LLM clusters recent fluid memories and produces abstractions with citations.
3. Store synthesized items with elevated phase hint ("stabilized") and dedicated kinds (`KIND_SYNTHESIZED_CONVENTION`, `KIND_PROJECT_PATTERN`, etc.).
6. Compaction / phase transitions proposed with dry-run + user review.
**cya Mapping:**
- Local JSON already provides a usable episodic stream.
-`export_memory` + recall with activation_context already return provenance.
- Orchestrator already injects memory into ContextEnvelope.
- Extend `retrospect` to offer synthesis as an explicit option.
- Add lightweight importance scoring on remember.
- Synthesized items flow through the same recall path (different kinds/phases).
**Delta Required:** Synthesis entrypoint (or call into phase-memory planner), new kinds, importance scoring, UI affordance in retrospect for "synthesize patterns".
- Structured objects with citation/provenance fields.
- Phase transition proposals (fluid → stabilized) with dry-run diffs.
- Multi-factor retrieval API that cya can parameterize.
**Safety / Explainability:** Every synthesized item carries machine + human readable citations. All influence is visible in `--explain-context`. Synthesis proposals are always reviewable (dry-run first).
**See also:** Research doc Variation 2, Park et al. (Generative Agents), MemoryVision phases section.
**Intent:** Allow the assistant to evolve its own high-level "how I should behave" rules and procedures as first-class, auditable, evolvable memory — the highest-leverage form of self-improvement, with the strongest guardrails.
**Core Loop:**
1. Base behavior driven by a dedicated tier of procedural memory items (`procedural_rule`, `meta_policy`, `explanation_strategy`, `safety_tuning`).
2. Meta-reflection (after retrospect or explicit "improve my rules"): LLM reviews recent outcomes + current rules + safety incidents and proposes patches.
3. Proposal + audit: phase-memory / cya presents structured diff ("+1 rule, safety impact: tightens") for user review/edit/approve/veto.
4. On approval: rule stored with high stability/phase and becomes active in future activation, prompt construction, and risk hints.
5. Guardrails: changes are additive or tightening only by default; every rule has provenance + last-review date; RiskClassifier treats procedural items as strong "force confirmation" signals.
**cya Mapping:**
- New dedicated kinds + `remember_procedural_rule` helper.
- New `cya improve-rules` (or retrospect extension) that triggers meta-reflection.
**Safety / Explainability:** Highest bar. All proposals declare safety impact. User veto is mandatory for anything that could relax posture. Full audit trail.
| Explainability | Full provenance + --explain-context | + actual reflection text | + citations on every synthesized item | + full audit trail on every rule change |
**Notes on the matrix:** Costs are relative. Profile 1 is designed to deliver quick wins on top of existing 0003 machinery. Profiles 2 and 3 require increasing collaboration with phase-memory planners and stronger dry-run/audit primitives.
### Handoff & Next Steps (within CYA-WP-0005)
- T04 turns the suggestions section of the research doc into the polished, standalone `docs/phase-memory-optimization-suggestions.md` (or equivalent) for sister-repo coordination.
- T05+ explore minimal implementation spikes (starting with Profile 1) only after the definitions and phase-memory feedback are reviewed.
- All profile work must preserve the Profile 0 invariants documented above.
See the full research artifact for deeper citations (Shinn, Park, LangMem, etc.) and the detailed phase-memory feedback list.
- Non-goals (this slice): full 9 kinds, embeddings/SemanticIndex, durable kontextual graph, voice, full profile authoring.
### Refined Port Signatures (cya seam)
These replace/extend the T05 no-op signatures. Implementations in T02+ will delegate to `phase_memory` (ports, planner, lifecycle, runtime or high-level sugar).
```python
def remember_preference(
key: str,
value: Any,
scope: str = "cwd",
*,
profile: str | None = None, # e.g. "cya-assistant-v1" or user profile id
ttl: str | None = None, # e.g. "30d" or phase hint
) -> None: ...
def recall_preferences(
scope: str = "cwd",
task_class: str | None = None,
*,
kinds: list[str] | None = None, # ["preference", "task"] etc.
# Includes status, phase info, provenance summary, policy notes for explain.
...
```
All calls must be non-blocking for the assistance path; failures → graceful empty + stderr warn (current behavior preserved).
### Ownership & Responsibilities
- **cya owns**: the seam (these 4 functions + wiring in orchestrator/cli for context + explain), safety integration (memory signals feed rule-based RiskClassifier but never bypass confirmation), user-visible explainability (provenance rendered in --explain-context and final output), graceful degradation.
Memory is now automatically activated based on the current working directory and git root. The orchestrator passes `activation_context`, and `recall_preferences` boosts relevant items. Activation is fully visible in `--explain-context` with provenance.
- **Retrospection as a First-Class Interaction Pattern (T04)**
Added the `cya retrospect` subcommand — a guided terminal reflection session. Users review recent memory usage, reflect, and explicitly record:
- Interaction goals
- Refined preferences
- Safety rules
These are stored with special kinds (`retrospection`, `interaction_goal`) and get preferential treatment in future activations.
- **Port & Data Model Extensions (T02)**
-`kind` parameter on `remember_preference`
-`activation_context` support on `recall_preferences`
-`remember_retrospection_outcome()` helper
- Improved `export_memory` with kind filtering and `by_kind` summary
- Full backward compatibility maintained.
- **Tests & Observability (T05)**
Comprehensive tests for activation boosting, retrospection records, provenance, and graceful degradation.
### Impact
A user can now:
- Teach `cya` once in a project and have it automatically remember those preferences.
- Run `cya retrospect` periodically to steer how the assistant behaves in the future.
- See exactly which memories influenced any response.
These features directly realize the "Personalized Console Helper" and "continuous optimization" vision from INTENT.md while staying within the explicit port seam.
**Next logical deepening:** When `phase-memory` exposes stable high-level activation + profile APIs, the local JSON implementation can be replaced while keeping the same user experience and `cya retrospect` flow.
This document is distinct from the Intent-vs-Scope gap analysis. It is the forward-looking vision for how memory will evolve in `cya` once real `phase-memory` integration begins. It should be updated as integration work progresses and as phase-memory itself matures.
