Graphify vs Solograph

Graphify (Apache-2.0 OR MIT) is the first tool in this space that overlaps Solograph hard enough to force a decision rather than a bookmark. Both parse code with tree-sitter into a graph an agent queries instead of grepping. They diverge on almost everything else — and several of Graphify’s choices are better than ours.

The one-line difference

Graphify is a deterministic, explainable graph of one corpus, shipped as a skill to 20+ assistants, with no embeddings anywhere. Solograph is a multi-project memory system — code graph plus episodic session history plus a semantic KB plus web search — shipped as one MCP server for Claude Code.

Graphify goes deeper on a repo. Solograph goes wider across a portfolio and remembers what you did.

Side by side

Their benchmark: LOCOMO (n=300) recall@10 of 0.497 against mem0 at 0.048 and supermemory at 0.149. Their harness, their numbers — untested here, and LOCOMO is a memory benchmark being used to argue about code retrieval, so read it as directional at best.

What to steal

Four mechanics are independently useful and don’t require adopting Graphify:

1. Tag every edge with provenance. EXTRACTED (literally in the source) vs INFERRED (resolved by the tool) turns “the graph says X calls Y” into a claim the agent can weigh. Our schema has IMPORTS / CALLS / INHERITS with no such distinction, so a heuristically-resolved dynamic dispatch looks exactly as solid as a literal import. Cheapest high-value change on this list.

2. PreToolUse hook on file reads. Graphify intercepts the agent before a Read, points it at a graph query, and in strict mode blocks the first raw read outright before reverting to a nudge. We already run PostToolUse hooks for reindexing — the same machinery, aimed one step earlier. This is harness engineering in its purest form: not “remember to query the graph” in CLAUDE.md, but a mechanism that makes ignoring it inconvenient.

3. Rebuild on git events, not on demand. post-commit + post-checkout means the graph is never stale and never needs a human to remember make scan. AST-only rebuilds cost nothing, so there’s no reason to defer them.

4. Tiered loading for explain — borrowed from OpenViking rather than Graphify, but it belongs in the same change: L0 abstract (~100 tokens) → L1 overview (~2k) → L2 full detail on demand. codegraph_explain currently returns one fixed-size answer; most calls only need L0 to decide whether to keep digging.

Plus one lower-priority idea: Leiden community detection to auto-partition a project into subsystems. That’s a better explain for an unfamiliar repo than a flat symbol dump.

What not to copy

• Dropping embeddings entirely. Graphify’s “no vector store” is a real advantage for code — ASTs are exact, embeddings drift. It’s the wrong call for our KB, where kb_search answers RU-language conceptual queries over prose that a graph can’t index. Keep the graph deterministic and the prose semantic; the mistake would be letting one purity argument override both.
• Sending docs and PDFs to a model. Graphify parses code locally but routes docs/PDFs/images through the assistant’s model. Fine for open source, wrong for 6-crm/ and anything under ~/personal/. Our privacy posture means non-code corpora stay local or don’t get indexed.
• Committing the graph. The merge driver is clever, but it exists because they commit graph.json. A local DB has no merge problem to solve — take the git hooks, skip the committed artifact.

Which to run

• One repo, unfamiliar codebase, want explainability and 36 languages → Graphify. It’s better at that specific job than our codegraph, and it costs nothing to try: uv tool install graphifyy && graphify install.
• Portfolio across ~/startups/active/, session history, RU+EN knowledge base, web search in the same tool → Solograph. Graphify has no equivalent of session_search, kb_search, or cross-project Cypher, and multi-project is the whole reason our graph exists.
• Both is coherent: Graphify per-repo for depth during a deep dive, Solograph as the portfolio-and-memory layer. They don’t share storage, so the cost is duplicated parsing, not conflict.

Caveats

Open core with a commercial upsell — “Graphify Enterprise” is on a waitlist at graphify.com as an always-on layer across meetings and files. The CLI itself is fully open with visible extraction code and no telemetry, so the downside risk is the usual one: the open core stops getting the interesting features. Note also the package-name trap: PyPI package is graphifyy (double y), CLI is graphify, and at least two lookalike domains serve the same pitch — verify you’re installing from the real repo.

Star counts are meaningless here. Three sources reported 76K, 97K, and 98K on the same day.

Links

• Graphify-Labs/graphify — the repo
• project-solograph — our code intelligence MCP server
• codegraph-guide — Solograph’s architecture and Cypher patterns
• agent-toolkit-landscape — where both sit in the ecosystem
• agent-memory-architecture — the 4-type memory frame both are instances of
• rag-patterns — graph retrieval vs the six other approaches
• agent-mistake-fix-harness — why the PreToolUse hook matters more than the graph format