Redis agent memory with Lettuce

Build a Redis-backed agent memory layer in Java with Lettuce, DJL (PyTorch), and standard Redis commands — working memory in a Hash, long-term semantic recall as JSON with a vector index, and an event log in a Stream.

This guide shows you how to build a small Redis-backed agent memory layer in Java with Lettuce and DJL (the Deep Java Library), using only standard Redis commands — no agent-memory SDK, no managed service. It includes a local web server built with the JDK's com.sun.net.httpserver so you can send turns at the agent, watch working memory update in place, see semantically similar long-term memories recalled in real time, watch the write-time deduplication skip near-duplicates, and inspect the per-thread event log.

The embedder is DJL (ai.djl.huggingface.tokenizers + ai.djl.pytorch.pytorch-model-zoo) running the canonical sentence-transformers/all-MiniLM-L6-v2 PyTorch checkpoint — the same library and model the existing Lettuce vector-search example uses, and the same encoder the Python example loads. DJL drives libtorch through the same C++ runtime as Python's PyTorch bindings, so the vectors produced here are numerically identical to the Python ones to within rounding noise, and the distance bands the Python walkthrough quotes carry over to this demo without recalibration. A memory written by one demo can be recalled by the other against the same Redis instance.

The big shape difference from the sister Jedis port is that Lettuce 6.7 doesn't ship first-class FT.* or JSON.* bindings yet. The helper classes send those commands through Lettuce's generic dispatch() API with a custom ProtocolKeyword, and the FT.SEARCH/FT.INFO replies are parsed by walking the resulting List<Object> — see LongTermMemory.java for the parser. The same connection is pinned to RESP2 at startup so the reply shape stays a flat array (Lettuce 6.7 negotiates RESP3 by default, which wraps the data in a map structure the parser would have to special-case).

Overview

The memory layer splits across three Redis primitives, each handling one tier:

  • Working memory for the active session is a Hash at agent:session:<threadId> holding the goal, scratchpad, a rolling window of recent turns (as a JSON list inside one field), and a few audit timestamps. One HGETALL returns the whole session in a single round trip; every write refreshes the key's EXPIRE so idle sessions decay on their own.
  • Long-term memory is a set of JSON documents at agent:mem:<id>, each carrying the memory text, a 384-dimensional embedding vector, and tag fields for user, namespace, kind (episodic / semantic), and source thread. A single Redis Search index covers the HNSW vector field and every metadata field, so one FT.SEARCH call performs the KNN with the metadata pre-filter in the same round trip. Write-time deduplication runs the same KNN at insert time and skips a new memory whose nearest existing entry is within a tighter threshold.
  • Event log for the agent's actions and observations is a Stream at agent:events:<threadId>, appended with XADD MAXLEN ~ so retention stays bounded automatically, replayed with XREVRANGE.

That gives you:

  • One Redis Search call per recall: FT.SEARCH does the KNN + TAG pre-filter in a single round trip (a per-row TTL follow-up is the only other read the helper issues, just to populate the ttl_seconds field for the admin panel). Working memory is one HGETALL; the event log is one XADD.
  • Sub-millisecond reads on every step of the agent loop, so the memory layer doesn't dominate per-step latency.
  • Per-tier decay: short TTLs on working memory, longer on episodic memories, no TTL on semantic memories. Combined with a database-level eviction policy (LFU is the common choice), memory stays bounded under pressure.
  • Scoping enforced inside the query: a recall query for user=alice will never see user=bob's memories, because the TAG filter goes into the same FT.SEARCH call as the KNN.

How it works

Each turn through the agent loop touches all three tiers in one pass: append to working memory, recall similar long-term memories, write the turn back as a new memory (with deduplication), and append one event to the log.

Per-turn flow

  1. The application calls embedder.encodeOne(text) to turn the incoming turn into a 384-element float[].
  2. session.appendTurn(threadId, role, content, user, agent, null) reads the per-thread Hash with HGETALL, appends the new turn to the rolling window in application code, trims it back to the configured maximum, and writes the Hash back with HSET + EXPIRE inside Lettuce's multi() / exec() transaction. The session TTL refreshes on every write so an active thread stays alive.
  3. memory.recall(vec, user, namespace, null, 5, threshold) issues FT.SEARCH via dispatch() with a TAG pre-filter and a KNN 5 clause. Redis returns the closest matching memories together with their cosine distances; memories beyond the recall threshold are dropped before they reach the agent so an unrelated query doesn't surface confident-looking false positives.
  4. memory.remember(text, vec, user, namespace, kind, threadId, null) runs the same KNN with a tighter dedup threshold. If an existing memory is within the threshold, the new write is skipped and the existing memory's hit_count is incremented with JSON.NUMINCRBY (also via dispatch()) — best-effort: if the memory's TTL has elapsed between the recall and the bump, the increment quietly fails and the hit count for that recall is lost. Otherwise a fresh JSON document is written with JSON.SET and a per-kind EXPIRE inside the same multi() transaction.
  5. events.record(threadId, action, detail) appends one entry to the per-thread Stream with XADD MAXLEN ~ (XAddArgs.Builder.maxlen(N).approximateTrimming()), bounding retention to roughly a thousand entries per thread without an explicit cleanup job.

The embedding is computed once and reused for steps 3 and 4 — there's no point encoding the same text twice. Recall runs before the write, so the agent doesn't see its own just-written turn echoed back as a recalled memory.

The session store

AgentSession wraps the working-memory Hash and the rolling turn window (source):

import com.redis.agentmem.AgentSession;
import com.redis.agentmem.SessionState;
import io.lettuce.core.ClientOptions;
import io.lettuce.core.RedisClient;
import io.lettuce.core.api.StatefulRedisConnection;
import io.lettuce.core.codec.ByteArrayCodec;
import io.lettuce.core.protocol.ProtocolVersion;

RedisClient client = RedisClient.create("redis://localhost:6379");
// Pin to RESP2 so FT.SEARCH / FT.INFO reply with the flat arrays
// the helper's parser expects (Lettuce 6.7 otherwise negotiates
// RESP3 against Redis 7+ and wraps the same data in a map shape —
// see the "Sending FT.* / JSON.* commands through Lettuce" section
// below for the longer story).
client.setOptions(ClientOptions.builder()
        .protocolVersion(ProtocolVersion.RESP2)
        .build());
StatefulRedisConnection<byte[], byte[]> connection = client.connect(ByteArrayCodec.INSTANCE);

// Shared lock across helpers so concurrent MULTI/EXEC spans on the
// single connection don't interleave queued commands.
Object txLock = new Object();
AgentSession session = new AgentSession(
        connection, txLock, "agent:session:", 3600, 20);

String threadId = session.newThreadId();
SessionState state = session.start(threadId, "alice", "demo-agent",
        "Plan next week's meetings.", null);
state = session.appendTurn(
        threadId, "user", "Schedule a budget review with finance.",
        "alice", "demo-agent", null);
System.out.println(state.turnCount() + " " + state.recentTurns().size()
        + " " + state.ttlSeconds());

The data model is one Hash per thread. The rolling turn window is stored as a JSON string in a single field so the whole session loads in one HGETALL — the hash never grows in size or field count as the conversation goes on.

agent:session:9f3d2a4b8c61
  thread_id=9f3d2a4b8c61
  user=alice
  agent=demo-agent
  goal=Plan next week's meetings.
  scratchpad=Need to confirm finance's availability.
  turn_count=4
  created_ts=1715990400.12
  last_active_ts=1715990650.83
  recent_turns=[{"role":"user","content":"...","ts":...}, ...]

Every write — start, appendTurn, setGoal — runs the HSET and EXPIRE inside multi() / exec() so a connection drop between the two writes can't leave the session without a TTL. The shared txLock serializes this whole MULTI…EXEC span against any other transaction on the connection — Lettuce's transaction state is connection-scoped, so two concurrent handler threads queueing into the same MULTI would interleave their writes.

The long-term memory store

LongTermMemory owns the JSON documents, the vector index, the recall query, and the write-time deduplication (source):

import com.redis.agentmem.LocalEmbedder;
import com.redis.agentmem.LongTermMemory;
import com.redis.agentmem.MemoryRecord;
import com.redis.agentmem.WriteResult;

LongTermMemory memory = new LongTermMemory(
        connection,
        txLock,
        "agentmem:idx",
        "agent:mem:",
        384,
        0.20,   // dedup threshold — tight at write time
        0.55,   // recall threshold — looser at read time
        null);  // default per-kind TTL map
LocalEmbedder embedder = LocalEmbedder.create();
memory.createIndex();  // idempotent

// Write a memory. The same KNN that powers recall also runs here at
// a tighter threshold so paraphrases of the same fact collapse.
float[] vec = embedder.encodeOne("The user prefers light mode in editors.");
WriteResult result = memory.remember(
        "The user prefers light mode in editors.",
        vec,
        "alice",
        "default",
        "semantic",
        "9f3d2a4b8c61",
        null);
System.out.printf("deduped=%s id=%s dist=%s%n",
        result.deduped(), result.id(), result.existingDistance());

// Recall against a later question.
float[] q = embedder.encodeOne("Which theme does this user like?");
for (MemoryRecord h : memory.recall(q, "alice", "default", null, 5, null)) {
    System.out.printf("%.3f [%s] %s%n", h.distance(), h.kind(), h.text());
}

Sending FT.* / JSON.* commands through Lettuce

Lettuce 6.7 doesn't ship first-class FT.* or JSON.* bindings, so the helper sends them through dispatch() with a custom ProtocolKeyword. Each command is its own keyword whose getBytes() returns the literal command name; the rest is built up with CommandArgs<byte[], byte[]>:

private enum ModuleCommand implements ProtocolKeyword {
    FT_CREATE("FT.CREATE"),
    FT_SEARCH("FT.SEARCH"),
    FT_INFO("FT.INFO"),
    FT_DROPINDEX("FT.DROPINDEX"),
    JSON_SET("JSON.SET"),
    JSON_NUMINCRBY("JSON.NUMINCRBY");
    // ...implementation elided
}

CommandArgs<byte[], byte[]> args = new CommandArgs<>(ByteArrayCodec.INSTANCE)
        .add(indexNameBytes)
        .add(filterClause + "=>[KNN 5 @embedding $vec AS distance]")
        .add("PARAMS").add(2).add("vec".getBytes(UTF_8)).add(vecBytes)
        .add("RETURN").add(8).add("user").add("namespace").add("kind")
        .add("source_thread").add("text").add("created_ts")
        .add("hit_count").add("distance")
        .add("SORTBY").add("distance").add("ASC")
        .add("LIMIT").add(0).add(5)
        .add("DIALECT").add(2);

List<Object> raw = sync.dispatch(
        ModuleCommand.FT_SEARCH,
        new NestedMultiOutput<>(ByteArrayCodec.INSTANCE),
        args);

The NestedMultiOutput collects the RESP reply into a List<Object> — a flat [count, key1, fields1, key2, fields2, ...] array under RESP2, where each fields is itself a nested list of alternating field/value pairs. The helper walks that with parseAllHits() and fieldsToMap() (both private to LongTermMemory).

The connection is pinned to RESP2 at startup with ClientOptions.builder().protocolVersion(ProtocolVersion.RESP2). Lettuce 6.7 otherwise negotiates RESP3 against Redis 7+, which wraps FT.SEARCH and FT.INFO results in map / set shapes that the flat-array parser would have to special-case across every reply path.

Data model

Each memory is a JSON document at agent:mem:<id>. The embedding is stored as a JSON array of floats so the document is human-readable from redis-cli; FT.SEARCH still expects the query vector as raw float32 bytes (LocalEmbedder.toBytes() packs them in little-endian order), regardless of how the indexed document stores it.

agent:mem:7c3f8a1b9e02
{
  "id": "7c3f8a1b9e02",
  "user": "alice",
  "namespace": "default",
  "kind": "semantic",
  "source_thread": "9f3d2a4b8c61",
  "text": "The user prefers light mode in editors.",
  "embedding": [0.013, -0.041, ...],
  "created_ts": 1715990400.12,
  "hit_count": 0
}

The Redis Search index is declared on the JSON document type with alias names on each path so the query syntax stays compact:

FT.CREATE agentmem:idx
  ON JSON PREFIX 1 agent:mem:
  SCHEMA
    $.text          AS text          TEXT
    $.user          AS user          TAG
    $.namespace     AS namespace     TAG
    $.kind          AS kind          TAG
    $.source_thread AS source_thread TAG
    $.created_ts    AS created_ts    NUMERIC SORTABLE
    $.hit_count     AS hit_count     NUMERIC SORTABLE
    $.embedding     AS embedding     VECTOR HNSW 6
                                       TYPE FLOAT32 DIM 384
                                       DISTANCE_METRIC COSINE

The query

Both recall and dedup share the same hybrid query: a TAG pre-filter in parentheses followed by =>[KNN k @embedding $vec]. With DIALECT 2, Redis applies the filter first and KNN-ranks only the matching documents.

FT.SEARCH agentmem:idx
  "(@user:{alice} @namespace:{default} @kind:{semantic})
     =>[KNN 5 @embedding $vec AS distance]"
  PARAMS 2 vec <384-float32-bytes>
  SORTBY distance
  RETURN 8 user namespace kind source_thread text created_ts hit_count distance
  DIALECT 2

distance is the cosine distance (0 means identical, 2 means opposite). Recall and dedup share the same query shape; only the threshold differs — strict at write time so the index doesn't fill with paraphrases of the same fact, looser at read time so the agent gets a wider net of relevant memories.

Per-kind TTLs

remember resolves the entry's TTL from the memory's kind:

Kind Default TTL When to use it
episodic 7 days Snapshots from a specific session that should decay.
semantic none Distilled facts and preferences the agent carries forward.

You can override per write by passing a non-null ttlSeconds to remember, or hand a different Map<String, Long> to the LongTermMemory constructor — for example, to give semantic memories a six-month TTL while leaving episodic memories at seven days.

The event log

AgentEventLog is a thin wrapper over a per-thread Redis Stream (source):

import com.redis.agentmem.AgentEvent;
import com.redis.agentmem.AgentEventLog;

AgentEventLog events = new AgentEventLog(connection, "agent:events:", 1000);
events.record(threadId, "turn_appended:user",
        "Schedule a budget review with finance.");
events.record(threadId, "memory_written",
        "wrote 7c3f8a1b9e02 as semantic");

for (AgentEvent e : events.recent(threadId, 20)) {
    System.out.println(e.action() + " " + e.detail());
}

record calls XADD with MAXLEN ~ 1000 via XAddArgs.Builder.maxlen(1000).approximateTrimming(). The tilde lets Redis trim in whole-node units instead of exactly-N units, which is much cheaper at the cost of overshooting the bound by up to a node's worth — the right tradeoff for an audit log where exact length doesn't matter.

The Stream is independent of the session Hash and the long-term JSON documents: it answers "what just happened" without competing with either of those for indexing or memory budget. Consumer groups (not used in this demo) would let downstream workers — summarisers, consolidators, audit pipelines — replay the log without losing position.

Concurrency caveats

The three helpers above trade correctness under heavy concurrency for clarity. Each is fine on a single-process demo, but lifting the code into a real multi-worker agent surfaces three races worth knowing about:

  • Working memory is read-modify-write. AgentSession.appendTurn calls HGETALL, mutates the recentTurns list in application code, and writes the Hash back with HSET. Two concurrent turns on the same thread can both read the same recentTurns, append different entries, and write back — last writer wins, the other turn is silently lost. The robust fix is either a WATCH / MULTI / EXEC loop around the read-modify-write or a small Lua script that does the append atomically server-side.

  • Long-term dedup is not atomic. LongTermMemory.remember runs a FT.SEARCH KNN lookup, decides whether the candidate is a duplicate, and (if not) calls JSON.SET. Two workers seeing the same fact in flight can each fail to see the other's not-yet-committed write and both insert a new memory. The pragmatic fix is to accept that the index will occasionally hold near-duplicates and run a background consolidator that periodically scans for memory pairs within a tight distance and merges them, rather than trying to make the write itself atomic.

  • The active thread is server state. The demo server keeps a single currentThreadId synchronized through an explicit mutex; seedAll, newThread, and handleTurn each release the lock between operations, so a turn racing with a thread rotation can capture the old id and apply to the previous thread. This is cosmetic for a one-user browser demo. A multi-user agent would carry the thread id on the request itself rather than as shared server state.

Two concerns specific to Java + Lettuce:

  • LocalEmbedder.encodeOne / encodeMany are synchronized because the underlying Predictor is not thread-safe. The demo's Executors.newCachedThreadPool could otherwise call into one predictor from several handler threads at once and corrupt the inference state. A higher-throughput deployment would replace that lock with a small pool of Predictor instances or a dedicated single-threaded inference executor.

  • All three helpers share a single Lettuce connection. Lettuce connections are thread-safe for individual command dispatch, but transaction state (MULTI / queued commands / EXEC) is connection-scoped — two concurrent threads queueing into the same MULTI would interleave their writes and each thread's EXEC would return a mix of replies. The shared txLock is passed into all three helpers (AgentSession, LongTermMemory, and AgentEventLog) and is held around every Redis call they issue, not only the MULTI…EXEC spans. That serializes every operation on the connection, which is fine for a single-user demo but trades concurrency for safety; a higher-throughput deployment would use a small pool of connections via Lettuce's ConnectionPoolSupport instead so handlers can run their commands in parallel.

Those caveats are deliberate. A more conservative implementation would obscure the Redis-shaped parts of the pattern; the demo prioritizes a small, readable code path that maps directly onto the commands in the prose above.

Pre-seeding long-term memory

In a real deployment the memory store fills up organically as the agent reasons over user turns: each turn produces zero or more memories that flow into the store, with deduplication catching repeats. For the demo, SeedMemory pre-loads a small set of mixed semantic and episodic memories so the very first recall query returns something useful (source):

import com.redis.agentmem.SeedMemory;

LongTermMemory memory = new LongTermMemory(connection, txLock,
        "agentmem:idx", "agent:mem:", 384, 0.20, 0.55, null);
LocalEmbedder embedder = LocalEmbedder.create();
memory.createIndex();
int written = SeedMemory.seed(memory, embedder, "default", "default", "seed");
System.out.println("seeded " + written + " memories");

The seed list mixes long-lived facts and preferences (semantic) with snapshots of past sessions (episodic), so the Kind to write control in the demo has something to switch between when a new turn is being remembered.

The interactive demo

DemoServer runs the JDK's HttpServer on port 8093, with a cached thread pool dispatching requests to handlers. The HTML page exposes three live panels — working memory, recalled memories, event log — plus a memories table for admin actions. Endpoints:

Endpoint What it does
GET /state Index info, current session, in-scope long-term memories, and recent events.
POST /turn Embed the text, append to working memory, recall similar memories, optionally write a new memory (with dedup), append an event.
POST /new_thread Start a fresh thread; long-term memory and other threads are untouched.
POST /reset Drop every long-term memory and re-seed the sample set.
POST /drop_memory Delete a single long-term memory by id.

The server holds one LocalEmbedder, one AgentSession, one LongTermMemory, and one AgentEventLog for the lifetime of the process. The "current thread" is a mutex-protected String field that the New thread button rotates — every browser tab inherits the same thread until you explicitly start a new one.

Run the demo locally

  1. Clone the redis/docs repository and change into the example directory:

    git clone https://github.com/redis/docs.git
    cd docs/content/develop/use-cases/agent-memory/java-lettuce
    
  2. Build the fat jar. You'll need a JDK 17 or later and Maven:

    mvn -q package
    

    The first build pulls Lettuce, DJL, and the PyTorch native libraries — that takes a couple of minutes the first time and is cached afterwards.

  3. Make sure a Redis instance with Redis Search and Redis JSON is running locally on port 6379. Redis Stack ships both, or Redis 8 with the Search and JSON modules enabled.

  4. Start the demo. The first run downloads the sentence-transformers/all-MiniLM-L6-v2 PyTorch weights into the local DJL cache (~90 MB):

    java -jar target/agent-memory-lettuce.jar
    

    Or via Maven: mvn -q exec:java.

  5. Open http://localhost:8093 and try some turns:

    • "Remind me which theme I prefer in editors." — paraphrase of a seeded semantic memory ("The user dislikes dark mode and prefers a high-contrast light theme..."). You should see that memory recalled with a cosine distance around 0.47, comfortably under the 0.55 default recall threshold.
    • "What did we discuss about the order-routing outage?" — paraphrase of a seeded episodic memory; the postmortem memory should recall around 0.44. Switch the Kind to write dropdown to skip so the question itself doesn't enter long-term memory.
    • "I prefer concise answers without filler phrases." — paraphrase of a seeded semantic memory. Switch the Kind to write dropdown to semantic so the dedup KNN runs in the same kind as the seed (dedup is scoped per kind, on purpose, so an episodic write can't collapse onto a semantic memory). You should then see the write deduped onto the existing memory at a cosine distance around 0.15, with hit_count ticking up in the memories table.
    • "My favorite color is teal." — unrelated to any seed; nothing recalls above the threshold (every seed lands above 0.8), and the new memory is written as episodic (or semantic, depending on the dropdown) under a fresh id.
    • Switch the User field to bob and re-ask any of the above — recall returns nothing because the seed memories live under default. That's the TAG pre-filter at work inside FT.SEARCH.
    • Slide the Recall threshold down to 0.30 to see borderline paraphrases drop out of the recall set, then back up to 0.70 to watch them return.

    DJL drives libtorch through the same C++ kernel as Python's PyTorch bindings, so distances here match the Python demo to four decimal places. sentence-transformers/all-MiniLM-L6-v2 puts a faithful paraphrase in the 0.15 – 0.50 cosine-distance range, a loose paraphrase or related topic in the 0.50 – 0.80 range, and unrelated queries above 0.8 — which is what motivates the 0.55 default recall threshold and the 0.20 default dedup threshold. A stricter embedding model (or a domain-tuned one) would let you tighten both; a noisier one would push them up. The right thresholds are always a function of the model, the corpus, and how conservative the agent needs to be about accepting a memory as a match.

The server is read/write against your local Redis. The default memory index is agentmem:idx, JSON keys live under agent:mem:, session Hashes under agent:session:, and event Streams under agent:events:. Useful flags (pass them after the jar):

  • --host / --port — change the HTTP bind address (default 127.0.0.1:8093).
  • --redis-host / --redis-port — point at a non-local Redis (default localhost:6379).
  • --mem-index-name / --mem-key-prefix / --session-key-prefix / --event-key-prefix — relocate the index name and the three key prefixes (to run several demos against one Redis without colliding, for example).
  • --no-reset — keep the existing long-term memories across restarts instead of dropping and re-seeding.
  • --session-ttl-seconds — change the working-memory TTL (default 3600).
  • --dedup-threshold — change the cosine-distance cutoff for write-time deduplication.
  • --recall-threshold — change the default cosine-distance cutoff for recall.
RATE THIS PAGE
Back to top ↑