memgram · now in public beta

Memory infrastructure for AI agents.

AI agents learn preferences, history, decisions, and context. Memgram turns those interactions into memory you can retrieve, inspect, and govern.

Backed bytechstars_NYC
app.memgraph.com — memory layer
live
conversation
We migrated off Salesforce four months ago — pilot is on Growth plan, 500 seats.
Logged. Routing future tickets to the self-hosted evaluation track.
Security review is still in progress, hold rollout until data residency is confirmed.
Got it — flagging deployment as blocked on EU residency sign-off.
memory pipeline
124ms
1
extract
ok
Security review required before rollout.
2
classify
ok
constraint · deployment.blocker
3
deduplicate
ok
supersedes prior 'review scheduled' note
4
decide
persisted
persist · scope=account.acme-enterprise
memory state
account · acme-enterprise3 active
constraint0.92
Security review required before rollout.
status0.84
Pilot expanded from 50 to 500 seats.
fact0.78
Migrated from Salesforce 4 months ago.
Scope: account.acme-enterprise
events
2,418
persisted
1,892
rejected
526
p50 trace
118ms
VPC deployableOpen core · coming soonOpenTelemetry compatible
01 · Problem

Memory gets harder over time.

The first few conversations usually work.

Then preferences change. Facts become outdated. Issues get resolved. New information conflicts with old information.

Most memory systems can store information. Very few help you understand what changed, what was retrieved, or why.

memory state · account.acme-enterprise14 weeks
  1. Pilot signed: 50 seats on Growth plan, self-hosted evaluation.
    Week 1
    stored as account status
  2. Security review required before any production rollout.
    Week 3
    constraint added — supersedes 'review scheduled'
  3. Pilot expanded from 50 to 500 seats.
    Week 6
    status changed — old seat count still retrieved
  4. Security review completed and signed off.
    Week 10
    old blocker still surfaces in retrieval
  5. Two memories now contradict each other on rollout readiness.
    Week 14
    no signal which one is current
Without inspection, you can't tell which memory the agent will use next.
02 · When you need memory

Do you actually need memory?

If all five are true, you're building a memory problem.

Common examples
  • Support agents
  • Healthcare assistants
  • Sales agents
  • Research systems
  • Legal workflows
  • Your agent works across multiple sessions
  • Your agent learns things from conversations
  • Information changes over time
  • Context doesn't naturally belong in a database
  • Wrong context creates bad outcomes
03 · Use cases

What can you build with memory?

Memory becomes valuable when agents learn from interactions and improve over time.

Healthcare
Smart Patient Care Assistant

Remembers patient history, preferences, and treatment context across visits.

Provides more personalized care without making patients repeat themselves.

Chronic Condition Companion

Learns what works, what doesn't, and how conditions evolve over time.

Delivers support tailored to each patient's journey.

Therapy Progress Tracker

Builds on previous sessions, goals, and emotional context.

Creates continuity and trust across conversations.

Education
Personalized Learning Assistant

Remembers strengths, weaknesses, and learning goals.

Adapts lessons as students progress.

Exam Preparation Coach

Tracks mastered concepts and recurring mistakes.

Focuses attention where it matters most.

Student Success Companion

Builds a long-term understanding of each learner.

Creates continuity across every session.

E-commerce
Shopping Assistant

Remembers preferences, interests, and purchase history.

Delivers recommendations that improve over time.

Personal Stylist

Learns favorite brands, styles, and sizing preferences.

Makes product discovery more relevant.

Customer Journey Assistant

Maintains context across browsing, purchasing, and support.

Creates a more seamless customer experience.

Customer Support
Technical Support Agent

Remembers previous issues and troubleshooting history.

Reduces repetitive questioning and speeds up resolution.

Account Support Assistant

Tracks customer preferences and recurring problems.

Provides support with full context from the start.

Customer Success Agent

Builds understanding across every interaction.

Helps teams deliver more proactive support.

Sales & CRM
Sales Development Agent

Remembers stakeholders, objections, and buying signals.

Makes follow-ups more relevant and contextual.

Account Intelligence Assistant

Builds a living understanding of every customer relationship.

Maintains continuity across long sales cycles.

Revenue Copilot

Tracks customer priorities and evolving requirements.

Helps teams make better decisions with better context.

Research & Analysis
Research Assistant

Remembers findings, hypotheses, decisions, and source context across investigations.

Helps teams build knowledge that compounds over time instead of restarting every research session.

Market Intelligence Agent

Learns sector trends, competitive moves, and signal patterns across reports.

Surfaces insights that emerge only from connecting many sources over time.

Due Diligence Copilot

Builds a cumulative picture of risks, findings, and open questions.

Reduces redundant analysis and improves decision quality.

If your agent learns information that changes over time and doesn't naturally belong in a database, you're dealing with a memory problem.

03 · Capabilities

Everything you need to run memory in production.

Cognition pipeline

What happens to a memory after it is captured?

Watch extraction, classification, deduplication, and persistence happen in real time.

Retrieval causality

Why was this memory retrieved?

Every retrieval comes with a reason. See what surfaced, what was scored, what was rejected.

Belief state inspector

What does the system currently believe about this user?

Open any user, session, or agent and see what the system currently believes.

Policy & governance

Who controls what persists and what expires?

Control what persists, what expires, what merges, and what stays isolated — per scope.

Multi-agent scopes

Can multiple agents share memory without leaking context?

Prevent cross-agent contamination. Inspect every transition between agents sharing a user.

Managed or self-hosted

Where does the memory infrastructure run?

Start on managed cloud. Self-hosted VPC coming soon — same APIs, same control plane.

02 · Why current memory systems break

Today's memory is a black box.

Memories get stored automatically, retrieved probabilistically, and you have no real answer to what your system believes or why.

Current systems
memory.store(?)
???
memory.recall(?)
  • stores memories automatically
  • retrieves probabilistically
  • hides extraction logic
  • no observability
memgram
app.memgraph.com — accounts / acme-enterprise
live
account
acme-enterprise
first seen 4 months ago · last active 14m ago
support-agent · 12sales-agent · 4
memory state
3 active
Scope: account.acme-enterprise — shared across support, sales, and CS
status
pilot expanded from 50 → 500 seats
0.92
constraint
security review required before rollout
0.88
incident
usage meter lag · inc_8821 open
0.74
activity — newest first
all agents ▾all events ▾
SEARCH3 results · 84ms14m
open blockers for acme-enterprise rollout
WRITEsupport-agent · 1 stored20h
Customer reports usage meter lag exceeding 15 minutes.
linked to inc_8821 · incident · 0.74
WRITEsales-agent · 1 rejected20h
Customer reports usage meter lag exceeding 15 minutes.
cross-scope write blocked · policy: support.write_only
  • every decision is a traceable event
  • retrieval comes with a reason
  • inspect what the system believes
  • policies you control
04 · The product

Inspectable runtime state for AI agents.

Trace every state transition. See what your agent currently believes. Replay any retrieval decision. Memgram gives you the visibility and control needed to run agent memory in production.

1 persisted

Why this memory was retained.

Every memory event opens into a full trace: the user message, extraction instructions, and every step of processing — extract, classify, deduplicate, decide.

  • step-by-step pipeline replay
  • per-event extraction reasoning
  • reportable, shareable trace IDs
app.memgraph.com — trace 5a565eef
live
Trace
5a565eef…
globex-health · support-agent · sess_3f7a… · 20h ago
1 persisted
user message
Our usage meter is lagging again — invoices are off by ~3%.
Reopening incident inc_8821 and attaching this report.
extraction instructions
Extract incidents, blockers, constraints, and account status changes.
Skip pleasantries, greetings, and acknowledgements.
memory processing
extract
42mscompleted
"Customer reports usage meter lag exceeding 15 minutes."
classify
18mscompleted
categorised as · incident · billing.metering
deduplicate
22mscompleted
linked to open incident inc_8821 — appending evidence
decide
11mscompleted
persist · scope=account.globex-health · confidence 0.84
app.memgraph.com — accounts / acme-enterprise
live
account
acme-enterprise
first seen 4 months ago · last active 14m ago
support-agent · 12sales-agent · 4
memory state
3 active
Scope: account.acme-enterprise — shared across support, sales, and CS
status
pilot expanded from 50 → 500 seats
0.92
constraint
security review required before rollout
0.88
incident
usage meter lag · inc_8821 open
0.74
activity — newest first
all agents ▾all events ▾
SEARCH3 results · 84ms14m
open blockers for acme-enterprise rollout
WRITEsupport-agent · 1 stored20h
Customer reports usage meter lag exceeding 15 minutes.
linked to inc_8821 · incident · 0.74
WRITEsales-agent · 1 rejected20h
Customer reports usage meter lag exceeding 15 minutes.
cross-scope write blocked · policy: support.write_only
Belief state

What the system currently believes.

Open any user, session, or agent to see the active belief state alongside every write, search, and rejection. Catch cross-agent contamination before it reaches production.

Playground

Watch memory evolve in real time.

Chat with an agent and see — turn by turn — what's extracted, what persists, what's retrieved, and what's rejected. Copy the integration straight into your stack.

Playground — watch memory work in real time
live
chat
We're on the Growth plan, self-hosted evaluation, and need feature flags on every release. No external API calls in staging — that broke us during the Redis outage on May 14.
Understood. Could you share the deployment target and which environments need the flag-gating enforced before rollout?
no memory context
Production deployment is blocked again — IAM permissions missing on the new role.
Opening an incident for the IAM gap. Given your staging policy and the May 14 Redis postmortem, I'll route this to platform-ops before any retry.
3 memories used
Type a message…
what's happening
↓ Session started650ms
0 memories for workspace_8471 · cold start
TURN 1
Search skipped · no memories yet
⚙ Processing complete
extracted
stack: Next.js + PostgreSQL
feature flags required for production
no external API calls in staging
TURN 2
search
1182ms
3 results
session memory
Next.js + PostgreSQL on Vercel + RDS.
stackturn 1
Feature flags required for every production release.
conventionturn 1
No external API calls allowed from staging environments.
constraintturn 1
retrieved in last turn
0.84No external API calls allowed in staging.
0.71Redis outage on May 14 blocked deployment.
0.62Feature flags required for production releases.
integration code
turn 2
1 · session
import memgram
 
session = memgram.Session(
agent_slug="coding-copilot",
user_id="workspace_8471",
)
 
2 · search memory
# before reply
memories = session.search(
query="deploy blocked — IAM missing on new role",
)
# → 3 results
 
3 · store async
# after reply
session.add(messages=[…])
# → 0 stored, 0 rejected
05 · How it works

One pipeline. Every decision, visible.

memgram sits between your application and your model runtime. Every memory event passes through a typed pipeline you can introspect and control.

01
AI App
your agent, copilot, or workflow
02
memgram SDK
drop-in for TS, Python, Go
03
Memory processing
extract · classify · dedupe · decide
04
Observable memory layer
state, traces, policies
05
AI runtime
consistent, inspectable behavior
// instrumentation is one line
import { memgraph } from '@memgraph/sdk'
await memgraph.record(conversation)
const ctx = await memgraph.recall({ scope: user.id })
07 · Built for real agents

Templates for the agents you actually ship.

Start from a memory profile tuned to your agent's job — support, personal assistant, coding copilot, sales, health. memgram configures extraction and retention automatically.

app.memgraph.com — new agent
live
new agent · step 1 of 3
What does this agent do?
We'll configure memory extraction automatically.
Support bot
tickets · resolutions · history
Personal assistant
preferences · routines · events
Coding assistant
stack · conventions · repos
Sales bot
leads · intents · objections
Health assistant
vitals · habits · medications
Custom
define your own extraction rules
preview: extracts tickets · resolutions · sentiment
Continue
08 · Production failures

The failures you can't debug today.

Every team running agents in production has hit these. Without an inspectable cognition layer, the root cause looks like a hallucination — but it's a state transition no one saw.

stale context

Support agent quoted last quarter's billing flow to an enterprise customer.

A B2B SaaS team rolled out an LLM support agent for acme-enterprise. Two weeks in, it confidently referenced a deprecated metered-billing flow that had been replaced after the May 14 release — the old memory was never invalidated.

memgram exposes
  • what was extracted from the deprecated runbook
  • why it persisted past the policy window
  • which retrievals scored it above the new billing doc
cross-agent contamination

Sales agent leaked an open support incident into an outbound email.

The support-agent and sales-agent shared the account scope without an isolation policy. The sales-agent retrieved an open incident memory (inc_8821, usage meter lag) during a renewal touchpoint — and personalised the outreach on it.

memgram exposes
  • which agent wrote which memory
  • why the cross-scope retrieval matched
  • the exact policy rule that should have blocked it
retrieval drift

Coding copilot stopped recommending the team's actual stack.

Over 6,000 turns the embedding space drifted. The copilot started suggesting generic Express patterns because team-specific conventions (Next.js + PostgreSQL, feature flags required) fell below the retrieval threshold.

memgram exposes
  • confidence decay over time per memory
  • rejected retrievals and their scores
  • state transitions that changed the agent's belief
06 · Deployment

Managed cloud today. Self-hosted soon.

Start on managed cloud with enterprise-grade security. Self-hosted VPC and open-core distribution are landing for design partners — same APIs, same control plane.

Open core · self-hosted · coming soon
api.memgram.io · control plane
managed · p50 118ms
$ curl -X POST https://api.memgram.io/memory/add \
-H "X-API-Key: mem_xxx..." -d '{agent_slug:"support-agent",account_id:"acme-enterprise",...}'
{ trace_id: "3a8f1b2c", status: "processing" }
$ curl https://api.memgram.io/trace/3a8f1b2c
persist · status · "Pilot expanded 50 → 500 seats" ✓ new
supersede · constraint · "Security review required" ↻ resolved
reject · context · low importance (0.18) ✗ drop
pipeline_timing · extract 3487ms · dedup 51ms · persist 124ms
$ curl -X POST https://api.memgram.io/memory/search \
-d '{query:"open rollout blockers", account_id:"acme-enterprise", limit:5}'
2 results · 84ms · hybrid graph + vector
09 · Teams in production

Used by teams building production AI systems.

Before memgram, debugging long-running conversations felt mostly probabilistic. We could see prompts and traces, but not how the system's memory state evolved over time. The retrieval traces changed that immediately.
Aria Vance
Core AI Team
We were chasing a hallucination for weeks. Turned out to be cross-agent contamination, a memory written by support surfacing in the sales agent. memgram showed us the exact transition.
Daniel Okafor
AI Product Engineer
We didn't need another vector database. We needed visibility into what our agents actually believed about users across sessions. memgram gave us a much clearer operational model for long-term memory.
Maya Bergstrom
Staff AI Engineer
10 · Pricing

Simple plans. Scale when your agents do.

Every plan includes the full memgram cognition stack — observability, retrieval traces, and graph memory. You only pay for the writes your agents persist.

Free
$0/month
10,000 memory adds/month
  • Unlimited searches
  • Unlimited agents
  • Unlimited users
  • Hybrid BM25 + vector search
  • Graph memory
  • Full observability
Start free
Starter
$29/month
50,000 memory adds/month
  • Unlimited searches
  • Unlimited agents
  • Unlimited users
  • Hybrid BM25 + vector search
  • Graph memory
  • Full observability
Get started
Most popular
Growth
$129/month
225,000 memory adds/month
  • Unlimited searches
  • Unlimited agents
  • Unlimited users
  • Hybrid BM25 + vector search
  • Graph memory
  • Full observability
Get started
Pro
$299/month
500,000 memory adds/month
  • Unlimited searches
  • Unlimited agents
  • Unlimited users
  • Hybrid BM25 + vector search
  • Graph memory
  • Full observability
Get started
Enterprise
Let's talk.

Unlimited scale, dedicated infrastructure, and policies tailored to your compliance posture. For teams running mission-critical agents in production.

Contact us
  • Unlimited memory adds
  • Unlimited searches, agents & users
  • Dedicated VPC / self-hosted
  • Advanced auth & audit logs
  • Custom policies & SLAs
  • Priority support & onboarding

Stop guessing what your AI remembers.

memgram is the control plane for AI cognition — inspectable state, traceable retrieval, governable persistence. Built for production AI systems.