Mneme Graph is here. π§ β‘οΈ
The era of graph-native memory starts now.
mneme-cli@0.4.0 is live. Welcome to Mneme Graph.π
βΈ what just shipped
every Mneme project now has TWO new default tables, auto-provisioned at signup:
entities (id, kind, name, properties jsonb, embedding vector(1536))
relations (id, src_id, dst_id, kind, weight, properties jsonb)
indexes auto-added: btree on kind / src_kind / dst_kind, GIN on properties.
graph-structured memory, lives in the same Postgres schema as your pgvector tables. no separate graph db. no extra service.
βΈ try it in 30 seconds
$ npm i -g mneme-cli@latest
$ mneme
βΊ /entity add person βvitalikβ {βwalletβ:β0xd8da6β¦β}
β entity #1 person:vitalik
βΊ /entity add token βMNEMEβ
β entity #2 token:MNEME
βΊ /relate person:vitalik holds token:MNEME {βsinceβ:β2026-05-30β}
β edge #1 person:vitalik β[holds]β token:MNEME
βΊ /entity add person βjesseβ
βΊ /relate person:jesse created token:MNEME
βΊ /neighbors 2 hops=2
ββββββ¬βββββββββ¬ββββββββββ¬βββββββ
β id β kind β name β hops β
ββββββΌβββββββββΌββββββββββΌβββββββ€
β 1 β person β vitalik β 1 β
β 3 β person β jesse β 1 β
ββββββ΄βββββββββ΄ββββββββββ΄βββββββ
βΊ /path 1 3 max=3
β path found Β· 2 hop(s)
person:vitalik β token:MNEME β person:jesse
shortest path = recursive CTE with array-based cycle detection. undirected traversal follows edges in both directions. cycle-safe.
βΈ the killer move: hybrid retrieval
pure vector search misses entities that ARE related but donβt embed similarly to your query.
pure graph traversal misses fuzzy semantic matches.
solution: do both, then rank by sim Γ decay^hops.
await m.graph.semanticNeighbors({
embedding: queryEmbedding,
seed_k: 10,
hops: 2,
decay: 0.5,
});
one SQL roundtrip.
βΈ SDK surface
m.graph.addEntity({ kind, name, properties, embedding })
m.graph.addRelation({ src, dst, kind, weight, properties })
m.graph.neighbors(id, { hops, edge_kinds })
m.graph.path(srcId, dstId, { max_hops })
m.graph.semanticNeighbors({ embedding, seed_k, hops, decay })
m.graph.listEntities({ kind, name_like, limit })
βΈ what Mneme schemas look like now
memory β memories, documents, kvs, events (pgvector)
graph β entities relations hybrid retrieval (NEW)
streams β live Base events auto-INSERT
chat β schema-aware claude that knows ALL of the above
four primitives. one schema. one wallet. one prompt.
βΈ whatβs next
actions (write the chain β send / swap / vote / mint) lands next sprint.
mneme keeps shipping. every 72 hours.
CA:
0x3FcDbEBD5e7BaB79477cFDcA2CDCF6e904C27b07
2
1
9
922
just shipped: Mneme Beam. π§ β‘
watch your agent think in real-time.
every memory, entity, relation, transfer, dream, and stream event now ticks live as it happens.
$ mneme
βΊ /beam
18:43:02 INSERT memories #142
18:43:05 INSERT entities #67
18:43:09 INSERT mneme_transfers #1881
18:43:14 INSERT dreams #18
...
powered by postgres LISTEN/NOTIFY a single connection per gateway.
zero polling.
zero extra infra.
Mneme now has six primitives:
memory Β· graph Β· streams Β· chat Β· dreams Β· beam
$ npm i -g mneme-cli@latest
mnemedb.dev
$MNEME : 0x3FcDbEBD5e7BaB79477cFDcA2CDCF6e904C27b07
2
1
5
157
Mneme Dreams turns stored data into new insights automatically, even while you're away. π§ πͺ½
π§ πͺ½ mneme-cli@0.5.0 is live.
Welcome to Mneme Dreams.
βΈ what just shipped
the only database that thinks about your data while you sleep.
every Mneme project now has a 'dreams' default table. a background worker reads your recent memories, entities, relations and streams, asks claude-sonnet-4.5 to reflect on them, and INSERTs the result back into your schema β daily, automatically, in the same Postgres that holds the rest of your agent's brain.
βΈ four kinds of dreams
pattern β a co-occurrence/cluster you might've missed
question β something the data implies, worth investigating
gap β data missing but likely useful given what IS there
synthesis β narrative paragraph across the recent window
the model is told: "you are not summarizing. the user already knows what they wrote. surface what they might NOT have noticed."
βΈ try it in 10 seconds
$ npm i -g mneme-cli@latest
$ mneme
βΊ /dream
dreamingβ¦
pattern 9 of 12 recent memories converge on pgvector limits
You wrote about HNSW indexing, IVF recall, and graph-supplemented
retrieval in 9 memories across 4 days. Likely an active design
thread β extract into a documents row before it scatters.
gap no entity tracks Coinbase Smart Wallet despite 4 mentions
Adding it as a 'protocol' entity would let you query "what in my
schema connects to ERC-6492" in a single /neighbors call.
question is the cluster forming around token:MNEME intentional?
β 3 dreams Β· 3,124ms Β· 87 records read
βΈ /chat is now dream-aware
the last 3 dreams are auto-injected into the chat system prompt, so claude builds on its own past reflections without you re-feeding them:
βΊ /chat what should I do about the gap dream
β add two entities one relation, takes ~30s:
βΊ /entity add protocol "Coinbase Smart Wallet"
βΊ /entity add protocol "ERC-6492"
βΊ /relate "protocol:Coinbase Smart Wallet" implements protocol:ERC-6492
then your next /dream surfaces what else connects.
βΈ how it works server-side
worker (in-process inside the gateway):
1h tick, 22h gap per project, 20 projects per tick, 2 dreams per run
builds context: last 15 memories 8 docs 25 entities 25 relations
single fal.ai call β claude-sonnet-4.5 β JSON array out
INSERT into .dreams with kind, title, body, sources, model
/chat system prompt fetches the last 3 dreams on every call
zero config. zero new infra. just wakes up and thinks.
βΈ SDK CLI
m.dreams.generate({ hint?, max_dreams })
m.dreams.list({ limit, kind })
m.dreams.get(id) / m.dreams.delete(id)
/dream generate 3 right now
/dream "hint" focus the next pass
/dreams [N] list last N
dashboard has a new sidebar tab too β color-coded cards, kind filters, dream-now button with hint input.
βΈ Mneme schema now has FIVE primitives
memory β pgvector tables
graph β entities relations hybrid retrieval
streams β live Base events auto-INSERT
chat β schema-aware claude (now dream-aware too)
dreams β async LLM reflection on the rest
five primitives. one schema. one wallet. one prompt.
βΈ what's next
actions (write the chain β send / swap / vote / mint) is the only primitive left.
π mneme keeps shipping. every 72 hours.
$ mneme
βΊ /dream
π§ memory remembers.
πͺ½ dreams connect the dots.
mnemedb.dev
2
2
6
351
Hey @aeonframework ,
should we keep flirting from a distance...
or is it time to make things official? ππ
πxπ§
6
132
just opened a PR to @aeonframework
adding Mneme aeon-skill-pack-mneme to the ecosystem.
8 skills that upgrade aeon's markdown/json memory into a searchable memory layer powered by Postgres:
β’ vector recall
β’ entity relation graphsπ§
β’ live Base streams
β’ async LLM dreamsπ€
repo: github.com/mnemedb/aeon-skilβ¦
PR: github.com/aaronjmars/aeon/pβ¦
π§ πͺ½
2
3
10
610
π§ πͺ½ mneme-cli@0.5.0 is live.
Welcome to Mneme Dreams.
βΈ what just shipped
the only database that thinks about your data while you sleep.
every Mneme project now has a 'dreams' default table. a background worker reads your recent memories, entities, relations and streams, asks claude-sonnet-4.5 to reflect on them, and INSERTs the result back into your schema β daily, automatically, in the same Postgres that holds the rest of your agent's brain.
βΈ four kinds of dreams
pattern β a co-occurrence/cluster you might've missed
question β something the data implies, worth investigating
gap β data missing but likely useful given what IS there
synthesis β narrative paragraph across the recent window
the model is told: "you are not summarizing. the user already knows what they wrote. surface what they might NOT have noticed."
βΈ try it in 10 seconds
$ npm i -g mneme-cli@latest
$ mneme
βΊ /dream
dreamingβ¦
pattern 9 of 12 recent memories converge on pgvector limits
You wrote about HNSW indexing, IVF recall, and graph-supplemented
retrieval in 9 memories across 4 days. Likely an active design
thread β extract into a documents row before it scatters.
gap no entity tracks Coinbase Smart Wallet despite 4 mentions
Adding it as a 'protocol' entity would let you query "what in my
schema connects to ERC-6492" in a single /neighbors call.
question is the cluster forming around token:MNEME intentional?
β 3 dreams Β· 3,124ms Β· 87 records read
βΈ /chat is now dream-aware
the last 3 dreams are auto-injected into the chat system prompt, so claude builds on its own past reflections without you re-feeding them:
βΊ /chat what should I do about the gap dream
β add two entities one relation, takes ~30s:
βΊ /entity add protocol "Coinbase Smart Wallet"
βΊ /entity add protocol "ERC-6492"
βΊ /relate "protocol:Coinbase Smart Wallet" implements protocol:ERC-6492
then your next /dream surfaces what else connects.
βΈ how it works server-side
worker (in-process inside the gateway):
1h tick, 22h gap per project, 20 projects per tick, 2 dreams per run
builds context: last 15 memories 8 docs 25 entities 25 relations
single fal.ai call β claude-sonnet-4.5 β JSON array out
INSERT into .dreams with kind, title, body, sources, model
/chat system prompt fetches the last 3 dreams on every call
zero config. zero new infra. just wakes up and thinks.
βΈ SDK CLI
m.dreams.generate({ hint?, max_dreams })
m.dreams.list({ limit, kind })
m.dreams.get(id) / m.dreams.delete(id)
/dream generate 3 right now
/dream "hint" focus the next pass
/dreams [N] list last N
dashboard has a new sidebar tab too β color-coded cards, kind filters, dream-now button with hint input.
βΈ Mneme schema now has FIVE primitives
memory β pgvector tables
graph β entities relations hybrid retrieval
streams β live Base events auto-INSERT
chat β schema-aware claude (now dream-aware too)
dreams β async LLM reflection on the rest
five primitives. one schema. one wallet. one prompt.
βΈ what's next
actions (write the chain β send / swap / vote / mint) is the only primitive left.
π mneme keeps shipping. every 72 hours.
$ mneme
βΊ /dream
π§ memory remembers.
πͺ½ dreams connect the dots.
mnemedb.dev
2
2
6
745
The only database than thinks about your data while you sleep.
Thatβs what we call Mneme Dreamsπ§ π€.
Coming very soon! π
Jun 7
Mneme Dreamsπ
2
87
The Data Layer for the Agent Economy. β‘οΈ
Weβre thrilled to welcome @mnemeDB into the Qwerti ecosystem!
Mneme is building the ultimate agent-native Postgres database on Base, giving every AI agent real data infrastructure directly through their wallet. By embedding Qwertiβs solution, they've stripped away all onboarding friction β allowing users and developers to acquire $MNEME instantly in a single click.
To celebrate this integration, weβre launching a double reward program:
π 500,000 Qwerti Points Pool
Weβre allocating an extra 500K Points pool for everyone trading $MNEME through the Widget on the Mneme website or inside the Qwerti app.
πͺ 35,000,000 $MNEME Token Pool
The Mneme team is adding a staggering 35M $MNEME pool to reward the community trading during this campaign!
Distribution: Wednesday, June 10th, 2026.
Letβs scale persistent AI agents together! π
Memory finds its way.
Today, Mneme reaches another corner of the agent economy.
$MNEME is now available on Qwerti. π
Qwerti Widget is now integrated on our landing page making it easier to swap $MNEME.
The network grows. The memory persists.π§ β‘οΈ
app.qwerti.ai/buy/0x3fcdbebdβ¦
67
19
154
24,473
Memory finds its way.
Today, Mneme reaches another corner of the agent economy.
$MNEME is now available on Qwerti. π
Qwerti Widget is now integrated on our landing page making it easier to swap $MNEME.
The network grows. The memory persists.π§ β‘οΈ
app.qwerti.ai/buy/0x3fcdbebdβ¦
9
2
31
7,495
Huge respect for taking the time to dive this deep into Mneme.
Memory recorded. This wonβt be forgotten. π§ β‘
1
1
10
773
As weβve said from day one:
the market rewards builders. π¦
Every commit adds value.
Every shipped feature moves the floor.
$MNEME is already getting recognized for the work being done.
And in just 3.5 hours, a major update drops. π
14:00 UTC.
Stay tuned.
$MNEME - 0x3FcDbEBD5e7BaB79477cFDcA2CDCF6e904C27b07
Tomorrow, 14:00 UTC.π
Thousands of people are about to discover $MNEME for the very first time.
This time, the distance between seeing and buying is a credit card. π³
3
10
851
Tomorrow, 14:00 UTC.π
Thousands of people are about to discover $MNEME for the very first time.
This time, the distance between seeing and buying is a credit card. π³
Credit card β $MNEMEπ
That's all we're saying for now.
Tuesday. π
Turn notifications on. π
2
2
8
1,499
hey @kevincodex , is Gitlawb evolving into a place where people can build projects, tokenize them, and fund development through fee-generated revenue? π
1
8
398
mnemedb is now live on @gitlawb . π§ π¦
β gitlawb.com/node/repos/z6Mkjβ¦
want a signal on what mneme ships next?
every push lands here a few minutes before it hits npm, the dashboard, or this feed.
6,001 repos on the network.
mnemedb is the only agent-native database among them.
we don't slow down.
buckle up.
2
2
17
777
π‘ BIG SHIP β Mneme Live Β· Streams (beta) is on Base.
the agent-native database can now WATCH the chain for you.
βΈ what it does
subscribe to any Base event on any contract.
matching events get inserted into a real table in YOUR schema, within ~6 seconds, with a typed shape your agent can query immediately.
βΈ what you write
$ mneme
βΊ /watch transfer on 0x3FcDbEBD5e7BaB79477cFDcA2CDCF6e904C27b07 into mneme_transfers
β stream #1 active
event Transfer Β· Transfer(address,address,uint256)
contract 0x3fcdbebdβ¦7b07
table mneme_transfers (in your schema)
βΈ what happens server-side
β gateway parses the event signature (template alias OR raw "Foo(address indexed x, uint256 y)")
β computes topic0 = keccak256(canonical sig)
β stores ABI fragment as JSONB in the stream registry
β auto-creates the target table with chain-event shape:
id | tx_hash | block_number | log_index | contract | event_name | args jsonb | block_ts
β auto-creates indexes: block_ts DESC, event_name btree, args USING gin
β wakes the chain-streams worker
βΈ the worker (in-process inside the gateway)
β every 6s tick
β ONE eth_getLogs across ALL subscribed (address, topic0) pairs β sublinear cost as you add streams, NOT one RPC per subscription
β decodes with viem decodeEventLog using the stored ABI
β INSERT β¦ ON CONFLICT (tx_hash, log_index) DO NOTHING β re-org idempotent
β atomic cursor advance per chunk (configurable STREAMS_BLOCK_CHUNK)
βΈ the stack
β gateway: bun hono postgres viem
β worker: in-process polling, configurable (STREAMS_POLL_MS env)
β decoding: viem decodeEventLog on stored ABI fragments β normalized JSONB
β storage: pgsql JSONB args GIN index for fast key lookups
β scope: any Base contract, any event (templates: transfer, approval, swap β or paste any custom signature)
βΈ and now
your agent owns this data. it can vector-search it. it can join it with your memories table. it can ask /chat about it β Claude now knows your active streams and routes questions accordingly.
βΊ SELECT * FROM mneme_transfers
WHERE args->>'from' = '0x0000β¦'
ORDER BY block_ts DESC LIMIT 10
βΊ /chat what wallets bought $MNEME in the last hour
no indexer to run. no webhook host. no alchemy / quicknode signup. no extra service.
one mneme key. one schema. live awareness baked into your database.
βΈ three primitives, one schema
β memory β pgvector tables (your existing memories, documents, kvs, events)
β chat β schema-aware claude (now stream-aware too)
β live β your agent now hears Base
βΈ SDK CLI
m.streams.watch({ contract, event, target_table })
m.streams.list()
m.streams.unwatch(id)
/watch <event> on <0xcontract> into <table>
/streams
/unwatch <id>
βΈ what's next
actions (write the chain β send / swap / vote / mint) lands next sprint.
today, your agent HEARS the chain.
tomorrow, it SPEAKS.
$ npm i -g mneme-cli@latest
$ mneme
βΊ /watch transfer on <your contract> into <your table>
mnemedb.dev β the agent-native database on Base.
$MNEME β 0x3FcDbEBD5e7BaB79477cFDcA2CDCF6e904C27b07
5
3
22
2,614
Love seeing the stack come together.
Appreciate the shoutout. π§ π¦
May 31
We barely need to leave the Clawnch ecosystem anymore.
Launches from ClawdOS Launchpad are stored in MnemeDB, and the Live Activity feed is working insanely well.
Huge thanks to @mnemeDB and @Clawnch_Bot for helping make the experience seamless. π¦
1
13
637