[Feature] Memory product roadmap: curation, recall, and research workflows

[fl4p]

Summary

Track the remaining product slices for local agent memory after the initial local store, recall, memory tools, remote privacy gates, review-first auto-capture, memory browser foundation, recalled-context visibility, and per-chat/project memory controls.

Current Status

Delivered on feat/local-agent-memory and deployed through integration/sim:

• Local memory store with project/global scopes.
• Memory tools for search, save, and forget.
• Local prompt recall for project and global memories.
• Review-first auto-capture candidates: chats draft suggestions, and nothing is used until the user saves it.
• Memory browser foundation for saved memories and pending suggestions.
• Pending suggestion approve, edit, and dismiss flows.
• Basic memory provenance and scope labels in the browser.
• Recalled memory visibility in assistant generation details.
• Per-chat and per-project memory on/off controls.
• Memory-off gates for recall, auto-capture, memory tools, prior memory tool history, and memory-derived compaction summaries.
• Remote-provider privacy scrub for memory-derived messages and tool history.
• Memory tool output sanitization and title/body length caps.
• Runtime allow-list guard so unavailable memory tools cannot execute from raw/structured/LiteRT tool calls.
• iOS simulator deploy verified for memory controls and memory copy polish.
• Android Pixel 8 deploy verified for memory controls; Android parity requested for the latest copy polish via integration/sim.

Problem / Motivation

The core memory plumbing now exists, but users doing personal research, laws, taxes, and hobbies need a trustworthy product experience around curation, provenance, scope, recall transparency, and long-term maintenance. Memory should feel local, inspectable, correctable, and safe by default.

Remaining Product Slices

Memory Browser And Curation UX

• Review pending suggestions before saving them.
• Edit a pending suggestion before approving it.
• Show source/provenance labels for manual, chat auto-capture, and assistant tool-saved memories.
• Show scope clearly: global vs project.
• Basic search across memories and pending suggestions.
• Basic scope filter for all/project/global.
• Edit existing saved memories.
• Add structured filters by tag, source type, jurisdiction, as-of date, project, and status.
• Show richer source details, including source excerpt and originating chat/message/document where available.
• Add merge/update flow when a suggestion duplicates an existing memory.
• Add bulk actions for approve, dismiss, delete, tag, and move scope.

Recall Transparency And Controls

• Show which memories were recalled for a response.
• Let users disable memory per chat or per project.
• Hide memory tools from chat/tool UI when memory is disabled.
• Scrub previous memory tool history when memory is disabled.
• Add a per-chat memory drawer/panel for inspecting active recall context.
• Add one-click "do not use this memory here" from recall UI.
• Add one-click "forget this" from recall UI.
• Explain why a memory was suggested or recalled in user-facing terms.

Retrieval Quality For Research/Law/Tax/Hobby Use Cases

• Improve ranking with project/global weighting, recency, importance, lexical+semantic blend, and duplicate suppression.
• Treat jurisdiction, as-of date, source type, and confidence as first-class ranking/filter signals.
• Format recalled memories with provenance and freshness warnings where relevant.
• Add test fixtures for laws, taxes, hobbies, and personal research workflows.
• Add performance tests for large local memory stores.

Capture Workflows Beyond Chat Auto-Capture

• Save selected chat text as a memory.
• Create pending auto-capture suggestions from local chats when enabled.
• Keep auto-capture review-first rather than silent-save.
• Save document/PDF/web snippets into memory with provenance.
• Add project-level capture defaults and tag presets.
• Tune auto-capture heuristics from real usage and add false-positive/false-negative tests.
• Keep sensitive-data detection conservative and easy to expand.

Privacy, Safety, And Trust Controls

• Keep remote-provider privacy tests as regression gates.
• Block memory recall and memory tools for remote generation.
• Scrub persisted memory tool messages before remote-provider payloads.
• Sanitize memory prompt/tool formatting against markup/control-token injection.
• Add export, import, backup, and delete-all memory controls.
• Add storage stats and per-project memory counts.
• Add audit/debug view for memory writes, deletes, and recalls.
• Add migration tests for memory tables and candidate queues.
• Consider local encryption or OS-protected storage for memory DBs.

Maintenance And Lifecycle

• Clean up orphaned project memories/candidates robustly when projects are deleted.
• Add archive/disable state for stale memories.
• Add periodic duplicate detection and merge suggestions.
• Add schema/version metadata for future migrations.

QA And Release Readiness

• Unit, integration, and RNTL coverage for the shipped memory foundation.
• Adversarial review completed for remote privacy, memory-off behavior, tool-call gating, and UI exposure.
• Verify memory controls on iOS simulator and Android Pixel 8.
• Add E2E coverage for approve/dismiss, manual save, recall, forget, project/global scope, and remote model switching.
• Add more adversarial tests for prompt injection in memory bodies and tool output.
• Verify behavior with no embedding model, failed embeddings, and very large memory bodies.

Alternatives Considered

• Keep memory tool-only with no curation UI: too opaque for high-trust personal research workflows.
• Silent auto-save: not acceptable for laws/taxes/personal data; review-first is safer and easier to trust.
• Remote-compatible memory sync: out of scope for the local-first product goal.

Platform

• Android
• iOS
• Both

Additional Context

Current foundation includes local memory storage, manual memory tools, local prompt recall, opt-in auto-capture candidates, pending suggestion review, project/global scopes, recalled-memory visibility, per-chat/project disable controls, and remote-provider scrubbing for memory-derived context.