Skip to content
Trove

Trove / Project

Roadmap

Open Trove work from the GitHub issue tracker, plus planned and aspirational features quoted from VISION.md, ENGINEERING.md, PRD.md and the release notes.

On this page
  1. First principles
  2. Sources
  3. Open issues
  4. Engine: memory hierarchy
  5. Engine: index and graph
  6. Engine: semantic layer and compression
  7. Engine: retrieval and reflection
  8. Engine: distribution
  9. Plugin
  10. Out of scope in the original requirements

Everything on this page is open, planned or aspirational; where main already covers part of an item, the item says so. Each item names its source: an open GitHub issue, or a document in the repository. For what the code does now, see How it works, Memory hierarchy and Retrieval levels.

First principles

Most open items trace back to a limit documented elsewhere in these docs. Retrieval is lexical, so it misses synonyms (#3). The SessionStart hook loads whole indexes, which grow with the trove (#1). Entries go stale and nothing marks them (#2). Agents without skills or hooks can only read files (#6). Reading the list this way shows which limit an item would lift.

Sources

VISION.md, ENGINEERING.md and PRD.md are now marked historical in the repository: they record the original vision and the v0.1 design and are no longer updated. Their items below are ideas, not commitments.

Open issues

From the issue tracker:

  • #24 macOS and Linux app for work memory. The README lists it as the one v0.3 track still to come: “one place for your work memory across projects and agents”.
  • #19 Trove v0.3: memory, semantic and context layer for coding agents. The epic; its other tracks are merged.
  • #1 Relevance-ranked, token-budgeted session loading.
  • #2 Entry staleness and validation (decay pass).
  • #3 Semantic recall (embeddings) beyond keyword grep.
  • #4 Passive capture suggestions at session boundaries.
  • #5 Re-inject the trove after compaction (PreCompact or compact hook).
  • #6 Expose the trove as a local Model Context Protocol (MCP) server. An open pull request, #39, adds a read-only server as trove mcp serve; it is not merged.
  • #7 Secret guard on write. Still open, though trove.sh guard and the refusal in trove remember now cover the core of it.
  • #8 trove.sh doctor to lint index integrity.
  • #9 trove.sh remove <slug> to delete an entry and its index line.
  • #10 --tag and --type filters for trove.sh grep.
  • #11 trove.sh stats summary command.

Engine: memory hierarchy

From VISION.md, for levels that the engine defines but does not build or retrieve:

L0, Working Memory:

Current task. Current edits. Current objective. Current reasoning.

Lifetime: minutes.

L6, Repository Memory:

Global knowledge. Glossary. Naming conventions. Coding standards. Folder conventions. Engineering philosophy. Never loaded unless needed.

A project trove’s conventions/ section now holds much of what L6 describes, but as files rendered into a skill, not as engine objects.

Engine: index and graph

From VISION.md:

Every repository is parsed. Store functions, classes, interfaces, enums, SQL, routes, migrations, models, commands, signals, workers, background jobs, and tests.

Every symbol knows: owner, dependencies, references, git history, complexity, embedding, hash.

Maintain imports, call graph, inheritance, database relationships, API relationships, background jobs, filesystem dependencies, and configuration dependencies.

From the v0.2.0 release notes, known limitations:

Dependency graph covers imports, not full call graphs

Engine: semantic layer and compression

From VISION.md:

Each symbol receives embedding, summary, keywords, intent, responsibilities, risk score, and examples. Do not embed raw files. Embed meaning.

Raw file → 400 token summary → 100 token summary → 20 token summary → 5 keyword fingerprint → embedding. Never lose provenance.

From the v0.2.0 release notes:

Module summaries are heuristic (not LLM-generated)

Semantic layer uses keyword overlap (embeddings slot ready)

trove ask added Okapi Best Matching 25 (BM25) ranking on main, but it is still lexical. The retrieval evaluation in evals/ exists so that a future ranking change can report its effect against a measured baseline.

Engine: retrieval and reflection

From VISION.md:

Summary → Signature → Implementation → Related symbols → Entire file → Neighbor files.

User Request → Intent Detection → Task Planner → Architecture Retrieval → Dependency Expansion → Semantic Retrieval → Symbol Expansion → Code Retrieval → Prompt Assembly → Model → Reflection → Memory Update.

After every task, generate lessons, new abstractions, architecture changes, mistakes, and engineering decisions. Persist automatically.

Every object stores last modified, git commit, hash, confidence, and staleness score. Only stale objects are rebuilt.

Continuously build glossary, subsystem map, ownership graph, data flow graph, API graph, database graph, service graph, queue graph, and scheduler graph. This becomes searchable.

The long-term goal, from VISION.md:

Trove should make a 200k context model behave like it has access to millions of tokens.

Engine: distribution

From the v0.2.0 release notes:

No pre-built binaries yet (build from source)

This is still true on main.

Plugin

From ENGINEERING.md, section 11, Future work:

prune / review skills for retiring stale entries.

An opt-in session-end capture hook that proposes entries from the transcript.

export / import to move a trove between machines as one bundle.

Recall ranking by tag/type and recency.

From PRD.md, “Could have (later)”:

Tag and type filtering in recall.

Optional richer auto-capture hook (opt-in) that proposes entries at session end.

The SessionEnd hook added in v0.2.0 only prints a reminder. It does not propose entries from the transcript.

Out of scope in the original requirements

From PRD.md, “Won’t have (for now)”:

Any network calls, telemetry, or hosted storage.

That line describes the v0.1 plugin. On main there is still no telemetry or hosted storage, but trove ask can call a model provider and trove ship can run a coding agent when you ask for it.