hotmemory

Agent memory as tables on Hotdata.

A memory record is a row in a managed table. The row carries the text, a scope, tags, source references, and the time span in which the fact was true. The table is an ordinary Hotdata table, so an agent can join its memory to its own data in one SQL query. None of the memory systems that we surveyed stores memory as typed columns in the same engine as the data of the consumer.

Status: alpha. A minor version before 1.0 can change a public surface, and CHANGELOG.md lists each change. The storage contract exists in Python with two drivers. MemoryStore runs in process memory. HotdataStore keeps records in one Hotdata managed database. The memory contract, Memory, runs over either driver.

The library has two layers:

• A storage contract. Put, get, list, search, and delete records in a namespace. Records are immutable. A new put under the same key creates a new revision.
• A memory contract. Remember facts, recall them inside a context budget, supersede a fact, forget by id or by horizon, and load the profile of a subject. Extraction from raw text is optional, and it takes an extractor callable that the caller supplies. load cuts a Markdown document into chunks, stores each chunk as an episode, and extracts facts that name their episode as a source. capture with an episode key stores its text as an episode in the same way, or as chunks with chunk_chars. profile returns facts, profiles, and procedures, never episodes. recall returns them too, and adds episodes after the facts only when the caller asks for them.

The library runs in the process of the consumer and calls the Hotdata API with the API key of the consumer. There is no hotmemory server. The first consumer is an incident investigator built on Hotdata. The library is not specific to it. A plain Python agent, a LangGraph agent, or any process with a Hotdata API key can use it.

Install

hotmemory needs Python 3.11 or later. The core package has no dependencies:

pip install hotmemory

Each driver and embedder that calls a service is in an extra. The hotdata extra adds HotdataStore. The openai extra adds OpenAIEmbedder.

pip install "hotmemory[hotdata,openai]"

Try the storage contract

MemoryStore needs an embedder for a search with query text. An embedder is a callable that turns a list of texts into a list of vectors. This example uses a toy embedder that counts two words:

from hotmemory import Filter, MemoryStore


def embed(texts):
    return [[text.count("disk") + 0.1, text.count("cpu") + 0.1] for text in texts]


store = MemoryStore(embedder=embed)
store.put(("team", "alerts"), "disk", kind="fact", content="The disk fills at night.")
store.put(("team", "alerts"), "disk", kind="fact", content="The disk fills at noon.")
store.put(("team", "alerts"), "cpu", kind="fact", content="The cpu spikes after a deploy.")

print(store.get(("team", "alerts"), "disk").id)  # team/alerts/disk@2
print([r.revision for r in store.history(("team", "alerts"), "disk")])  # [1, 2]
hits = store.search("disk", [("team",)], Filter(kind="fact"), k=1)
print(hits[0].record.content)  # The disk fills at noon.

Try the memory contract

Memory gives an agent the memory operations over any store. remember derives each key from the subject and the content, so a retried call writes nothing new. recall returns the records and one block of text, with one line for each record:

from hotmemory import Fact, Memory, MemoryStore


def embed(texts):
    return [[text.count("disk") + 0.1, text.count("cpu") + 0.1] for text in texts]


memory = Memory(MemoryStore(embedder=embed))
scope = ("team", "alerts")
memory.remember(
    [Fact(kind="fact", subject="disk", content="The disk fills at night.", sources=("chat-1",))],
    scope,
)
records, block = memory.recall("disk", [scope])
print(block)  # - The disk fills at night. [sources: chat-1] [valid: unknown to now]

load takes a document name, the Markdown text, a scope, and an extractor. The extractor gets each chunk, with the first heading and the heading path of the chunk at its top:

from hotmemory import Fact, Memory, MemoryStore


def embed(texts):
    return [[text.count("disk") + 0.1, text.count("cpu") + 0.1] for text in texts]


def last_line(text, observed_at, current):
    return [Fact(kind="fact", subject="disk", content=text.splitlines()[-1])]


memory = Memory(MemoryStore(embedder=embed))
text = "# Disk incident\n\n## Root cause\n\nThe disk fills at night.\n"
episodes, facts = memory.load("disk-incident", text, ("team", "alerts"), last_line)
print(episodes)  # ['team/alerts/disk-incident-0001@1']
episode = memory.store.get(("team", "alerts"), "disk-incident-0001")
print(episode.content.splitlines()[0])  # # Disk incident

Use the skill

skills/hotmemory/SKILL.md gives an agent the same operations as commands. The skill is a folder with the skill file and one script for each operation. The package does not include it, so you copy it from the repository.

1. Install hotmemory in the Python that the agent runs. The extras are necessary only for a Hotdata database:

   pip install "hotmemory[hotdata,openai]"
   

2. Copy the skill folder into the folder where your coding agent reads skills. Many agents use ~/.<agent>/skills/ for every project, or .<agent>/skills/ in one project. The documentation of your agent gives the correct folder.

   git clone --depth 1 https://github.com/hotdata-dev/hotmemory
   cp -r hotmemory/skills/hotmemory <skills-folder-of-your-agent>/
   

The agent runs the scripts from the skill folder. With --memory-file, they need no network and no key. With --database, they read the Hotdata and OpenAI keys from the environment.

Store memory in Hotdata

HotdataStore needs the hotdata extra, and an embedder for every write and search. The openai extra ships OpenAIEmbedder, which reads OPENAI_API_KEY. Any callable that turns a list of texts into a list of vectors also works. provision opens the database with the name, or creates it. It reads the connection from HOTDATA_API_KEY, HOTDATA_WORKSPACE, and HOTDATA_API_URL:

from hotmemory.hotdata import HotdataStore
from hotmemory.openai import OpenAIEmbedder

embedder = OpenAIEmbedder()
store = HotdataStore.provision(
    "agent-memory", embedder=embedder, model=embedder.model, dimensions=1536
)
store.put(("team", "alerts"), "disk", kind="fact", content="The disk fills at night.")
print([hit.record.id for hit in store.search("disk", [("team",)])])

One process writes to a database. docs/contracts.md gives the tables, the retrieval query, and the rules of the driver.

Benchmark

hotmemory ran LongMemEval-S, a public benchmark of 500 questions over long chat histories, on 2026-10-10. docs/benchmarks.md gives every number with its configuration.

On all 500 questions, on 2026-10-22, the memory arm scored 439 of 500 (87.8%). It ran capture with chunks of 1,000 characters and a gpt-4o-mini extractor, and recall with a budget of 16,000 characters, about 4,350 tokens of reader input. The reader and the judge were gpt-4o-2024-08-06. This is one run, and its 95% confidence interval is about 2.9 points on each side. The 392 questions that no tuning used scored 88.0%.

Retrieval needs no model except the embedder. The runner stored each session of a question as an episode in HotdataStore and searched with the question. The 419 scored questions had these results with the vector ranking. For 98.3% of them, an evidence session was in the top 5. For 99.3% of them, every evidence session was in the top 10.

Before that run, the end-to-end score ran on two samples of 54 questions, sample A on 2026-10-10 and sample B on 2026-10-14. Later rows give their own date. The reader and the judge were gpt-4o-2024-08-06.

Facts alone score below the full history. With the baseline prompt, the extractor did not store the answer in 19 of the 35 misses on sample A, and in 16 of the 27 misses on sample B. With the detail prompt, these counts are 16 of 23 and 14 of 22. With the whole sessions that a search adds after the facts, the extractor did not store the answer in 5 misses on each sample. With chunks of 1,000 characters, this count is 2 on each sample. The chunk row scores above the full history on both samples. It scores 3 more than the 10 sessions on sample A and 2 less on sample B, with about one seventh of their tokens. With 54 questions, each score can be about 10 points off. With the baseline prompt, the memory arm differed by 14.8 points between the two samples. In the cloud, a search took 361 ms at p50 with the vector ranking, and 998 ms with the fused ranking.

Documents

• docs/contracts.md: the record, the storage operations, the memory operations, and the platform facts behind them.
• docs/guarantees.md: each behavior that a consumer can rely on, its state, and its proof.
• docs/benchmarks.md: the numbers of hotmemory on LongMemEval-S, each with its configuration. bench/README.md tells you how to run it.
• CONTRIBUTING.md: the one command that checks a change, and the rules of the test suite.
• CHANGELOG.md: each change to a public surface.
• docs/architecture.md: how each operation runs on the Hotdata platform.