Skip to content

roadmap.md

Latest commit

 

History

History
130 lines (90 loc) · 4.81 KB

roadmap.md

File metadata and controls

130 lines (90 loc) · 4.81 KB

Product Roadmap

This page reflects the current reality of agent-memory, not the historical sequence that got the project here.

The project is no longer just a repository memory bootstrapper. It is now a broader memory platform with durable state, retrieval, automation, integration, retention, and self-host dogfooding capabilities.

Current Product Surface

Today agent-memory includes these product areas:

  • Canonical memory init, update, recall, status, validate
  • Retrieval query with natural-language modes, citations, and JSON output
  • External history ingestion add, sync
  • Local automation automate start|stop|status|run-once|ensure-running
  • Client integration integrate for Claude Code and Codex
  • MCP access mcp
  • Retention and archive management archive-first pruning under .agent-memory/archive/
  • Workflow layer higher-level workflow tools exposed through the MCP surface
  • Self-host dogfood tooling npm run dogfood:init|exercise|repair|status

What Is Already Complete

These foundations are established in the current codebase:

  • Durable repository memory with canonical state, history, checkpoints, sources, and config
  • Section-aware recall with policy controls and conservative deduplication
  • Query retrieval across bundle, history, and checkpoints with citations and evidence-insufficient handling
  • Status inspection with backlog, source health, checkpoint drift, and suggested next action
  • Retention and archive-first pruning integrated into the automation path
  • Local automation daemon for sync + recall maintenance
  • Claude Code and Codex integration surfaces
  • MCP tools and higher-level workflow entrypoints
  • Self-host dogfood loops for exercising and repairing the project against itself

Current Strategic Focus

The project has moved beyond the original Phase 1/2/3 framing. The meaningful frontier is now inside the broader platform.

The current focus areas are:

  • make the workflow layer the default operator experience
  • improve integration reliability for Claude Code and Codex
  • deepen automation and retention safety
  • expand the dogfood loop so the project can prove its own value continuously
  • keep the active memory surface small while the platform around it grows

Next Major Goals

1. Workflow-First Experience

Shift the product center of gravity from low-level commands to high-level workflows.

Near-term goals:

  • strengthen workflow tools such as memory_assess, memory_compact_handoff, and memory_maintain
  • make workflow outputs easier to trust and act on than raw command output
  • keep low-level commands available, but treat them as expert-mode tools

2. Integration Maturity

Turn integration from “available” into “boring and reliable.”

Near-term goals:

  • reduce mismatch states between managed integration assets and real project state
  • improve repair flows for partially broken Claude/Codex integration
  • tighten status/read-only inspection so operators can understand integration health without trial-and-error

3. Automation And Retention Safety

Automation now exists; the next step is making it resilient and predictable.

Near-term goals:

  • improve daemon observability and run diagnostics
  • harden archive-first pruning and archive expiry behavior
  • ensure active query/recall/status semantics stay clean even as archives grow
  • refine when aggressive auto-apply recall is acceptable versus risky

4. Broader Importer Coverage

Current importer support is intentionally narrow and local-first.

Near-term goals:

  • support more real-world Claude/Codex local history layouts
  • improve partial-failure handling and diagnostics
  • expand beyond current built-ins only when reliability and operator clarity are preserved

5. Dogfood And Repair Loop

The dogfood layer is now a strategic asset, not just internal tooling.

Near-term goals:

  • make the self-host exercise loop a stronger regression detector
  • improve deterministic repair before escalating to provider-driven repair
  • keep the dogfood worktree isolated while preserving realistic operator conditions

Longer-Horizon Opportunities

These are intentionally not immediate commitments, but they are plausible extensions of the current platform:

  • richer exporter/report surfaces for CI and dashboards
  • stronger policy controls for automation and retention
  • broader importer ecosystem
  • deeper MCP workflow composition
  • more agent-oriented output modes and structured handoff surfaces

Product Principle

The platform should continue to optimize for this balance:

  • keep active canonical memory small
  • keep historical evidence durable
  • make retrieval explainable
  • make maintenance deliberate
  • make automation observable
  • make integrations safe by default

That is the line between “a memory bootstrap tool” and “a trustworthy memory operating system for coding workflows.”