entergram (CLI)

Persistent engineering memory for AI coding agents. Cells are plain Markdown (source of truth); the index is a rebuildable cache. Zero native dependencies.

Install

npm install -g get-entergram        # once published
# or one-liner:
curl -fsSL https://entergram.dev/install | bash

Local dev (from this repo, before npm publish):

cd cli && npm install && npm link   # symlinks the `get-entergram` binary globally

Commands

get-entergram init                       # scaffold .entergram/ in the current repo
get-entergram learn [--source all|git|docs|pr|issue] [--limit 50] [--no-llm]
                                  # ingest git history + docs/README/ADRs + merged PRs → cells
get-entergram remember "<fact>" \        # write a fact by hand
        --why "…" --tags a,b --scope services/auth --type decision
get-entergram recall "<query>" [--budget 2000]   # FTS/bm25 retrieval under a token budget
get-entergram review                     # triage low-confidence auto-extracted cells
                                  #   --accept <id> | --reject <id> | --tag <id> --to a,b | --why <id> --to "…"
get-entergram reindex                    # rebuild the SQLite index from the cells
get-entergram doctor                     # health check (ids, hooks, dangling links, index coverage)
get-entergram serve                      # MCP server (stdio): recall/remember/learn/doctor as tools

The index is SQLite + FTS5 (.entergram/index.db, git-ignored and rebuildable). Markdown cells remain the source of truth.

Extraction (git/PR/issue → {decision, reason, outcome})

• Set a model key and learn extracts high-confidence structured memories from merged PRs/issues. Provider auto-selects: GEMINI_API_KEY (default, ENTERGRAM_GEMINI_MODEL default gemini-2.0-flash) or ANTHROPIC_API_KEY (ENTERGRAM_MODEL default claude-haiku-4-5). Force with ENTERGRAM_PROVIDER=gemini|anthropic|none.
• With no key, it falls back to a heuristic (title→what, body→why, confidence 0.4) so learn always works offline. Pass --no-llm to force heuristic.
• Requires gh installed + authenticated for PR/issue sources.

Connect your agent (MCP)

get-entergram serve exposes memory as MCP tools over stdio. Add it once and any MCP client uses it.

Claude Code — .mcp.json in your repo:

{ "mcpServers": { "get-entergram": { "command": "get-entergram", "args": ["serve"] } } }

Cursor / Windsurf — add an MCP server: command get-entergram, args ["serve"]. Continue — add the same block to config.json → mcpServers.

Tools: recall(query, budget?), remember(what, why?, tags?, scope?, type?), learn(source?, limit?), doctor().

What a cell looks like

---
id: B-0007
type: decision
tags: [auth, jwt]
scope: services/auth
source: {"kind":"commit","sha":"9f3a1c2","author":"@alice","date":"2026-05-14"}
confidence: 0.4
created: 2026-05-14
hook: Moved auth from sessions to JWT to kill sticky-session infra
---

# Moved auth from sessions to JWT to kill sticky-session infra
## What …  ## Why …  ## Outcome …

Roadmap (from the build spec)

• ✅ Day 4: SQLite/FTS index, docs/README/ADR ingestion, get-entergram review, get-entergram reindex.
• ✅ Day 5: get-entergram serve — MCP server (recall/remember/learn/doctor over stdio).
• ✅ Day 6: GitHub PR/issue harvest + LLM extraction of {decision, reason, outcome}.