ENGRAM is a multi-tenant database kernel built for the way LLMs actually think. Branch the world, mutate freely, merge what worked. Every write is provenance-stamped and shape-coerced against an ontology you control — exposed to your agent over MCP.
Traditional databases punish exploration. ENGRAM treats every speculative thought as a first-class object: forkable, mergeable, and forever attributable to the agent that authored it.
Spin up an ephemeral sandbox-* branch off @main in microseconds.
Agents reason, mutate, and explore in isolation — then merge what worked or
evaporate what didn't. Zero pollution of canonical state.
Define canonical fields once. ENGRAM uses vector boundaries to probabilistically
coerce noisy LLM output — annual_rev, revenue, and
yearly_sales all land in the same column.
Every write carries the agent ID, prompt hash, confidence score, and an optional C2PA manifest. The full causal chain of every fact is queryable, auditable, and cryptographically anchored.
The same multi-tenant RocksDB kernel speaks two protocols: a streaming MCP endpoint for agents, and a REST control plane for the humans who supervise them.
Bind a session to a tenant and call six tools: fork_world,
mutate_sandbox, get_entity, execute_cbs_search,
evaporate_sandbox, merge_sandbox. Returns are stable JSON
envelopes with accepted_fields / rejected_fields contracts.
Inspect the lineage tree, define ontology fields, and audit every mutation any agent has made. Three views: Multiverse Map, Ontology Registry, Provenance Explorer.
Point any MCP-aware client at https://api.inengram.ai/v1/spaces/default/mcp.
Below are working snippets — verified against the live v0.4.0 backend — for
Claude Desktop and a standalone Python MCP client. Copy, paste, and watch your agent write
to a forked branch in real time. The default space is a shared public sandbox;
for an isolated tenant, swap default for any slug you own.
{ "mcpServers": { "engram": { "command": "npx", "args": [ "-y", "mcp-remote", "https://api.inengram.ai/v1/spaces/default/mcp" ] } } }
# Connect, fork a sandbox, write a fact, read it back. import asyncio, json from mcp import ClientSession from mcp.client.streamable_http import streamablehttp_client ENGRAM = "https://api.inengram.ai/v1/spaces/default/mcp" async def main(): async with streamablehttp_client(ENGRAM) as (r, w, _): async with ClientSession(r, w) as s: await s.initialize() # 1. Branch off @main fork = await s.call_tool("fork_world", {"parent": "main"}) branch = json.loads(fork.content[0].text)["data"]["branch_id"] print("branched →", branch) # 2. Write a fact with full provenance (agent_id, prompt_hash, confidence_score) mut = await s.call_tool("mutate_sandbox", { "branch_id": branch, "node_id": "Acme Corp", "text_desc": "B2B SaaS", "payload": { "name": "Acme Corp", "founded_year": 2017, "annual_revenue": "$48M", }, "provenance": { "agent_id": "demo-agent", "prompt_hash": "sha256:demo", "confidence_score": 0.91, }, }) print(json.loads(mut.content[0].text)["data"]) # → {"written": true, "accepted_fields": ["name","founded_year","annual_revenue"], ...} # 3. Read it back from the forked branch got = await s.call_tool("get_entity", { "branch_id": branch, "node_id": "Acme Corp", }) print(json.loads(got.content[0].text)["data"]["payload"]) # → {"name": "Acme Corp", "founded_year": 2017, "annual_revenue": "$48M"} asyncio.run(main())
# Health curl -s https://api.inengram.ai/healthz # {"status":"ok","version":"0.4.0"} # Inspect the space's branch tree (every fork_world call appears here) curl -s https://api.inengram.ai/api/v1/spaces/default/lineage | jq # Pull the audit trail of every agent write curl -s https://api.inengram.ai/api/v1/spaces/default/provenance | jq
While your agent is writing, open a second terminal and run
curl -s https://api.inengram.ai/api/v1/spaces/default/lineage — every
fork_world call appears as a new branch; every
mutate_sandbox shows up in
/provenance with the agent ID and confidence score. The Developer
Console (coming soon) renders the same data as a live Multiverse Map.