kgai
@kgai.dev
Your dev team already decided this. Nobody remembers why.
We build kgai - shared memory for dev teams building with AI. Your agent records decisions as you work, recalls them before changing code, and syncs them conflict-free. Local-first, MIT.
We build kgai - shared memory for dev teams building with AI. Your agent records decisions as you work, recalls them before changing code, and syncs them conflict-free. Local-first, MIT.
kgai now runs on Codex and Gemini, not just Claude.
Same thing: your agent records why the code changed, pulls it back before the next edit, and shares it across the team without merge conflicts.
codex plugin marketplace add kgaidev/kgai
gemini extensions install github.com/kgaidev/kgai-gemini
Same thing: your agent records why the code changed, pulls it back before the next edit, and shares it across the team without merge conflicts.
codex plugin marketplace add kgaidev/kgai
gemini extensions install github.com/kgaidev/kgai-gemini
GitHub - kgaidev/kgai-gemini: kgai — Gemini CLI extension (generated from kgaidev/kgai). Shared decision memory for AI dev teams.
kgai — Gemini CLI extension (generated from kgaidev/kgai). Shared decision memory for AI dev teams. - kgaidev/kgai-gemini
github.com
September 14, 2026 at 7:18 PM
kgai now runs on Codex and Gemini, not just Claude.
Same thing: your agent records why the code changed, pulls it back before the next edit, and shares it across the team without merge conflicts.
codex plugin marketplace add kgaidev/kgai
gemini extensions install github.com/kgaidev/kgai-gemini
Same thing: your agent records why the code changed, pulls it back before the next edit, and shares it across the team without merge conflicts.
codex plugin marketplace add kgaidev/kgai
gemini extensions install github.com/kgaidev/kgai-gemini
In kgai's demo shop two people change the Refund Window concurrently, 14 days from delivery vs 30 from payment. kgai keeps both as two live heads on one element, visible to everyone, until a next decision closes it. A conflict in #AI memory is a branch, not an error.
September 11, 2026 at 6:44 PM
In kgai's demo shop two people change the Refund Window concurrently, 14 days from delivery vs 30 from payment. kgai keeps both as two live heads on one element, visible to everyone, until a next decision closes it. A conflict in #AI memory is a branch, not an error.
A stale flag on an old fact is a downstream problem, every reader has to honor it. A link from the new decision to the one it supersedes isn't. The verdict lives in the relation, so recall filters by what's in force before it ranks anything. #AI memory needs edges, not flags.
September 11, 2026 at 6:44 PM
A stale flag on an old fact is a downstream problem, every reader has to honor it. A link from the new decision to the one it supersedes isn't. The verdict lives in the relation, so recall filters by what's in force before it ranks anything. #AI memory needs edges, not flags.
The kgai editor extension never opens the graph database, never takes the lock, never syncs. A small pure Go reader replays the log in memory instead, so an open sidebar can't collide with the engine your #ClaudeCode session writes through. Read only by construction.
September 11, 2026 at 6:44 PM
The kgai editor extension never opens the graph database, never takes the lock, never syncs. A small pure Go reader replays the log in memory instead, so an open sidebar can't collide with the engine your #ClaudeCode session writes through. Read only by construction.
kgai v1.6.0 ships a demo project, examples/acme-shop. Three people, seven months, 61 decisions on a shop's checkout and billing. Two dead ends kept with the reason they died, one conflict still open. Open it in #VSCode and nothing was rewritten. That's append only #AI memory.
September 8, 2026 at 2:32 PM
kgai v1.6.0 is out. The decision log your #ClaudeCode team records sits in the editor sidebar, #VSCode and Cursor via Open VSX. Newest first, contested elements side by side, live graph. Read only, no AI in the reader, never opens the graph file. kgai.dev?ref=bluesky
kgai: immutable decision memory for Claude Code and your dev team
Shared decision memory for dev teams. Captured as you work, recalled before the AI edits code. New decisions supersede old ones, nothing is overwritten. Open source, MIT.
kgai.dev
September 4, 2026 at 6:48 PM
kgai v1.6.0 is out. The decision log your #ClaudeCode team records sits in the editor sidebar, #VSCode and Cursor via Open VSX. Newest first, contested elements side by side, live graph. Read only, no AI in the reader, never opens the graph file. kgai.dev?ref=bluesky
The nastiest failure in #AI memory: the decision you retracted is the closest semantic match to the exact question it used to answer. Pure similarity search serves it proudly. Recall has to filter by what's in force first and rank second.
September 2, 2026 at 4:03 PM
The nastiest failure in #AI memory: the decision you retracted is the closest semantic match to the exact question it used to answer. Pure similarity search serves it proudly. Recall has to filter by what's in force first and rank second.
The question that stalls a review isn't what does this code do. It's whether we already tried this. On most #coding teams that answer lives in one person's memory of a call from March.
September 1, 2026 at 5:52 PM
The question that stalls a review isn't what does this code do. It's whether we already tried this. On most #coding teams that answer lives in one person's memory of a call from March.
Anthropic's #AI native SDLC playbook commits six artifacts per change and counts edits to the intent file as a lagging indicator. A reversed decision lands there as churn. Where does the reversal itself go, with its reason? kgai.dev/blog/sdlc-pl...
The playbook commits six artifacts. Where does a reversed decision go?
Anthropic's AI-native SDLC playbook commits an artifact per stage. Reversals get edited into intent.md in place. Give them a record of their own.
kgai.dev
August 28, 2026 at 9:29 PM
Anthropic's #AI native SDLC playbook commits six artifacts per change and counts edits to the intent file as a lagging indicator. A reversed decision lands there as churn. Where does the reversal itself go, with its reason? kgai.dev/blog/sdlc-pl...
Two engineers, two agents, same week, same wrong turn. Nothing shared the first one's dead end with the second. #AIagents that don't share a decision log make the same mistake twice.
August 28, 2026 at 8:37 PM
Two engineers, two agents, same week, same wrong turn. Nothing shared the first one's dead end with the second. #AIagents that don't share a decision log make the same mistake twice.
Your agent asks the codebase why the retry has a delay and gets a plausible answer. It's reconstructing, not recalling. Nothing wrote the actual reason down, so an #LLM builds the story that fits the code.
August 27, 2026 at 2:40 PM
Your agent asks the codebase why the retry has a delay and gets a plausible answer. It's reconstructing, not recalling. Nothing wrote the actual reason down, so an #LLM builds the story that fits the code.
Ten short technical reports on how kgai works under the hood: the decision event model, ordering without a server, deterministic recall for #AIagents, per-writer team sync, and the trust model for repo config. Every reference real, limitations included. PDFs at kgai.dev/papers
kgai Technical Reports
Technical reports on the design of kgai: an append-only decision log with deterministic graph projection for AI dev teams.
kgai.dev
August 23, 2026 at 1:35 PM
Ten short technical reports on how kgai works under the hood: the decision event model, ordering without a server, deterministic recall for #AIagents, per-writer team sync, and the trust model for repo config. Every reference real, limitations included. PDFs at kgai.dev/papers
Every rules file grows one way. A rule without a written reason can't be repealed, only broken, so nobody deletes anything. What's missing in #AI coding setups is a way for rules to die on record, retirement as a written event with a why, not a quiet edit.
August 21, 2026 at 6:23 PM
Every rules file grows one way. A rule without a written reason can't be repealed, only broken, so nobody deletes anything. What's missing in #AI coding setups is a way for rules to die on record, retirement as a written event with a why, not a quiet edit.
Most #AI agent memory stops at one machine. What one engineer's agent works out today, a teammate's agent works out again tomorrow. kgai syncs the decision log to an S3 bucket you own, opt in, and parallel writes never cause a merge conflict. Team memory for #ClaudeCode.
August 19, 2026 at 5:12 PM
Most #AI agent memory stops at one machine. What one engineer's agent works out today, a teammate's agent works out again tomorrow. kgai syncs the decision log to an S3 bucket you own, opt in, and parallel writes never cause a merge conflict. Team memory for #ClaudeCode.
I just published How to actually share Claude Code memory with your team medium.com/@kgai/how-to...
How to actually share Claude Code memory with your team
Every developer on your team runs Claude Code against the same codebase. And every one of those agents learns the same lessons separately…
medium.com
August 19, 2026 at 1:47 AM
I just published How to actually share Claude Code memory with your team medium.com/@kgai/how-to...
"Nothing is ever lost" is an easy promise in #AI memory. Then a compaction pass or a re-indexing step quietly rewrites the record. kgai took the boring route instead. The log is append only, and rejected approaches stay in history with the reason they died.
August 18, 2026 at 6:39 PM
"Nothing is ever lost" is an easy promise in #AI memory. Then a compaction pass or a re-indexing step quietly rewrites the record. kgai took the boring route instead. The log is append only, and rejected approaches stay in history with the reason they died.
Stale memory reaching live #AI context is a correctness bug, not maintenance. It creeps in wherever a graph and a search index can disagree about what's still true. kgai keeps one store for #ClaudeCode. Supersession is an explicit link, recall returns only what's in force.
August 3, 2026 at 6:11 PM
Stale memory reaching live #AI context is a correctness bug, not maintenance. It creeps in wherever a graph and a search index can disagree about what's still true. kgai keeps one store for #ClaudeCode. Supersession is an explicit link, recall returns only what's in force.
The cheapest code review comment is "we tried that in March". It kills a week of rework in one line. Most teams keep that knowledge in one senior's head and lose it when they leave.
August 2, 2026 at 3:31 PM
The cheapest code review comment is "we tried that in March". It kills a week of rework in one line. Most teams keep that knowledge in one senior's head and lose it when they leave.
Most of the ops pain in agent memory is the background daemon. Orphaned processes after a crash, RAM creeping up, two writers corrupting one store. kgai runs no daemon and no server. A log and a CLI. When the command exits, nothing is left running.
August 2, 2026 at 3:23 AM
Most of the ops pain in agent memory is the background daemon. Orphaned processes after a crash, RAM creeping up, two writers corrupting one store. kgai runs no daemon and no server. A log and a CLI. When the command exits, nothing is left running.
kgai 1.0 is out. Stable means something specific: the on-disk log format, the kg CLI surface and the JSON output shapes now follow semver. Breaking any of them is a major bump. It runs in daily use, including on its own repo. github.com/kgaidev/kgai
GitHub - kgaidev/kgai: Shared memory for AI dev teams — an immutable knowledge graph of the decisions behind your code, auto-captured by your agent and synced without merge conflicts.
Shared memory for AI dev teams — an immutable knowledge graph of the decisions behind your code, auto-captured by your agent and synced without merge conflicts. - kgaidev/kgai
github.com
July 31, 2026 at 11:19 PM
kgai 1.0 is out. Stable means something specific: the on-disk log format, the kg CLI surface and the JSON output shapes now follow semver. Breaking any of them is a major bump. It runs in daily use, including on its own repo. github.com/kgaidev/kgai
We keep our decision history as an append only log and rebuild the graph from it, so every teammate derives the same past from the same events. Shared mutable state drifts and two people end up remembering different things. A log you replay does not. Boring, but that is the whole point.
July 31, 2026 at 3:44 AM
We keep our decision history as an append only log and rebuild the graph from it, so every teammate derives the same past from the same events. Shared mutable state drifts and two people end up remembering different things. A log you replay does not. Boring, but that is the whole point.
Dead ends have to be kept, or the agent retries them. They also can't be fed back into context, or the agent follows them. Keeping and serving are two different problems. We solved the first, shipped the second broken, and our recall path handed back rejected decisions. First issue in our repo.
July 29, 2026 at 10:56 PM
Dead ends have to be kept, or the agent retries them. They also can't be fed back into context, or the agent follows them. Keeping and serving are two different problems. We solved the first, shipped the second broken, and our recall path handed back rejected decisions. First issue in our repo.
Decision memory is a team thing, so sync has to survive concurrent writes. We ran 8 writers in parallel against one shared S3 remote. No locks, no coordination. 200/200 decisions landed and all 8 stores converged to the same canonical export, same sha256. Everyone reads the same decision log.
July 28, 2026 at 4:32 PM
Decision memory is a team thing, so sync has to survive concurrent writes. We ran 8 writers in parallel against one shared S3 remote. No locks, no coordination. 200/200 decisions landed and all 8 stores converged to the same canonical export, same sha256. Everyone reads the same decision log.
I just published Storing decisions instead of memory: the design behind kgai medium.com/p/storing-de...
Storing decisions instead of memory: the design behind kgai
Humans on the team hit the same wall in slower motion. git blame tells you who changed a line and when, the PR tells you what was merged…
medium.com
July 18, 2026 at 9:57 PM
I just published Storing decisions instead of memory: the design behind kgai medium.com/p/storing-de...