Literally the title. Ever sat in front of your computer screen, watching Claude burn tokens reading shit it has already read over and over again? Burning tokens that could have been used to write code or fix bugs?
That, my good friend, is exactly what Engram fixes. Engram is a shared context intelligence system that lets you reduce your token burn rate by a factor of huge magnitudes.
We measured this on a real codebase (42 TypeScript files, ranging from 1 KB to 42 KB per file).
For a task like "fix the billing page," an agent without Engram reads roughly 5 large files to orient itself, the kind that run 6,000–7,000 tokens each. That's around 31,500 tokens before writing a single line. With
Engram, get_context_for_task returns the ranked list of relevant symbols across all 42 files in about 2,700 tokens. The agent then reads the one file it actually needs around 2,800 tokens for a large component. Total:
around 5,500 tokens. That's roughly 6× cheaper, and the agent goes straight to the right file instead of
guessing.
For individual symbol lookups, get_symbol returns a function's signature, callers, callees, and test associations in about 95 tokens. A small utility file in this codebase costs around 450 tokens to read in full, whereas a large component is 6,000+. If an agent needs 10 symbol lookups across 10 small files, that's 950 tokens with
Engram vs 4,500 without. About 4× cheaper at the low end, and the savings compound sharply as files get larger.
For a targeted task like "fix the querySymbol function," the agent normally reads 4 core files, which adds up to around 26,000 tokens. With Engram, the agent makes 3 tool calls returning just the exact definitions, callers,
and blast radius, for about 320 tokens. That's roughly 80× cheaper for the same task-level understanding.
Every. Single. Session.
It even knows what changed since your last session so after a git pull, instead of rereading everything again, your agent gets a delta: "these 14 symbols changed, here's what's new." Cold starts become warm starts.
It even shows you the blast radius of a change before you make it.
get_dependencies("UserService", depth: 3)
tells you exactly what breaks. No more surprise regressions at runtime.
It even works across agents. Claude Code, Cursor, Windsurf, Cline, Zed, Gemini CLI. One MCP server, everything speaks to it.
The free plan is now officially ready, no credit card required. It runs 100% locally on your code which never leaves your machine. The paid tiers are for cloud sync and team graph sharing, which you only need when you have teammates and want to share and sync codebase context across agents across teams, and across computers.
go to the website and sign up for your free account now and stop wasting those tokens!
-Refraction-•1h ago
That, my good friend, is exactly what Engram fixes. Engram is a shared context intelligence system that lets you reduce your token burn rate by a factor of huge magnitudes.
We measured this on a real codebase (42 TypeScript files, ranging from 1 KB to 42 KB per file).
For a task like "fix the billing page," an agent without Engram reads roughly 5 large files to orient itself, the kind that run 6,000–7,000 tokens each. That's around 31,500 tokens before writing a single line. With Engram, get_context_for_task returns the ranked list of relevant symbols across all 42 files in about 2,700 tokens. The agent then reads the one file it actually needs around 2,800 tokens for a large component. Total: around 5,500 tokens. That's roughly 6× cheaper, and the agent goes straight to the right file instead of guessing.
For individual symbol lookups, get_symbol returns a function's signature, callers, callees, and test associations in about 95 tokens. A small utility file in this codebase costs around 450 tokens to read in full, whereas a large component is 6,000+. If an agent needs 10 symbol lookups across 10 small files, that's 950 tokens with Engram vs 4,500 without. About 4× cheaper at the low end, and the savings compound sharply as files get larger.
For a targeted task like "fix the querySymbol function," the agent normally reads 4 core files, which adds up to around 26,000 tokens. With Engram, the agent makes 3 tool calls returning just the exact definitions, callers, and blast radius, for about 320 tokens. That's roughly 80× cheaper for the same task-level understanding.
Every. Single. Session.
It even knows what changed since your last session so after a git pull, instead of rereading everything again, your agent gets a delta: "these 14 symbols changed, here's what's new." Cold starts become warm starts.
It even shows you the blast radius of a change before you make it.
get_dependencies("UserService", depth: 3) tells you exactly what breaks. No more surprise regressions at runtime.
It even works across agents. Claude Code, Cursor, Windsurf, Cline, Zed, Gemini CLI. One MCP server, everything speaks to it.
The free plan is now officially ready, no credit card required. It runs 100% locally on your code which never leaves your machine. The paid tiers are for cloud sync and team graph sharing, which you only need when you have teammates and want to share and sync codebase context across agents across teams, and across computers.
go to the website and sign up for your free account now and stop wasting those tokens!
That's it.